ollama source for Momentry Core verification

2026-05-22 17:19:10 +08:00
commit 0b31ff9135
2020 changed files with 1413145 additions and 0 deletions
--- a/x/models/glm4_moe_lite/glm4_moe_lite.go
+++ b/x/models/glm4_moe_lite/glm4_moe_lite.go
@@ -0,0 +1,770 @@
+// Package glm4_moe_lite provides the GLM4-MoE-Lite implementation for MLX.
+// This model uses Multi-head Latent Attention (MLA) and Mixture of Experts (MoE).
+package glm4_moe_lite
+
+import (
+	"encoding/json"
+	"fmt"
+	"math"
+
+	"github.com/ollama/ollama/x/mlxrunner/batch"
+	"github.com/ollama/ollama/x/mlxrunner/cache"
+	"github.com/ollama/ollama/x/mlxrunner/mlx"
+	"github.com/ollama/ollama/x/mlxrunner/model"
+	"github.com/ollama/ollama/x/mlxrunner/model/base"
+	"github.com/ollama/ollama/x/models/nn"
+	"github.com/ollama/ollama/x/tokenizer"
+)
+
+func init() {
+	base.Register("Glm4MoeLiteForCausalLM", newModel)
+	base.Register("GLM4MoeLite", newModel)
+}
+
+// RopeScaling holds RoPE scaling configuration
+type RopeScaling struct {
+	Factor       float32 `json:"factor"`
+	MscaleAllDim float32 `json:"mscale_all_dim"`
+}
+
+// Config holds GLM4-MoE-Lite model configuration
+type Config struct {
+	HiddenSize            int32   `json:"hidden_size"`
+	NumHiddenLayers       int32   `json:"num_hidden_layers"`
+	IntermediateSize      int32   `json:"intermediate_size"`
+	MoEIntermediateSize   int32   `json:"moe_intermediate_size"`
+	NumAttentionHeads     int32   `json:"num_attention_heads"`
+	NumKeyValueHeads      int32   `json:"num_key_value_heads"`
+	VocabSize             int32   `json:"vocab_size"`
+	RMSNormEps            float32 `json:"rms_norm_eps"`
+	RopeTheta             float32 `json:"rope_theta"`
+	MaxPositionEmbeddings int32   `json:"max_position_embeddings"`
+	AttentionBias         bool    `json:"attention_bias"`
+
+	// MLA (Multi-head Latent Attention) parameters
+	QLoraRank     int32 `json:"q_lora_rank"`
+	KVLoraRank    int32 `json:"kv_lora_rank"`
+	QKRopeHeadDim int32 `json:"qk_rope_head_dim"`
+	QKNopeHeadDim int32 `json:"qk_nope_head_dim"`
+	VHeadDim      int32 `json:"v_head_dim"`
+
+	// MoE parameters
+	NRoutedExperts      int32   `json:"n_routed_experts"`
+	NSharedExperts      int32   `json:"n_shared_experts"`
+	NumExpertsPerTok    int32   `json:"num_experts_per_tok"`
+	RoutedScalingFactor float32 `json:"routed_scaling_factor"`
+	NormTopKProb        bool    `json:"norm_topk_prob"`
+	FirstKDenseReplace  int32   `json:"first_k_dense_replace"`
+	NGroup              int32   `json:"n_group"`
+	TopKGroup           int32   `json:"topk_group"`
+
+	// RoPE scaling
+	RopeScaling *RopeScaling `json:"rope_scaling"`
+
+	// Quantization parameters (set during load based on model quantization)
+	QuantGroupSize int                               `json:"-"` // Group size for quantization (default 64)
+	QuantBits      int                               `json:"-"` // Bits per weight (4 or 8)
+	QuantMode      string                            `json:"-"` // Quantization mode ("affine", etc.)
+	TensorQuant    map[string]*model.TensorQuantInfo `json:"-"`
+
+	// Computed fields
+	QHeadDim int32   `json:"-"` // qk_nope_head_dim + qk_rope_head_dim
+	Scale    float32 `json:"-"` // 1/sqrt(QHeadDim) with mscale adjustment
+}
+
+// MLAAttention implements Multi-head Latent Attention with absorption.
+type MLAAttention struct {
+	QAProj      nn.LinearLayer
+	QALayerNorm *nn.RMSNorm
+	QBProj      nn.LinearLayer
+
+	KVAProjWithMQA nn.LinearLayer
+	KVALayerNorm   *nn.RMSNorm
+
+	EmbedQ     *nn.MultiLinear
+	UnembedOut *nn.MultiLinear
+
+	OProj nn.LinearLayer
+}
+
+// Forward computes absorbed MLA attention output.
+func (a *MLAAttention) Forward(x *mlx.Array, b *batch.Batch, c cache.Cache, positions *mlx.Array, B, L int32, cfg *Config) *mlx.Array {
+	q := a.QAProj.Forward(x)
+	q = a.QALayerNorm.Forward(q, cfg.RMSNormEps)
+	q = a.QBProj.Forward(q)
+
+	q = mlx.Reshape(q, B, L, cfg.NumAttentionHeads, cfg.QHeadDim)
+	q = mlx.Transpose(q, 0, 2, 1, 3)
+
+	qNope := mlx.SliceStartStop(q, []int32{0, 0, 0, 0}, []int32{B, cfg.NumAttentionHeads, L, cfg.QKNopeHeadDim})
+	qPE := mlx.SliceStartStop(q, []int32{0, 0, 0, cfg.QKNopeHeadDim}, []int32{B, cfg.NumAttentionHeads, L, cfg.QHeadDim})
+
+	compressedKV := a.KVAProjWithMQA.Forward(x)
+
+	kvCompressed := mlx.SliceStartStop(compressedKV, []int32{0, 0, 0}, []int32{B, L, cfg.KVLoraRank})
+	kPE := mlx.SliceStartStop(compressedKV, []int32{0, 0, cfg.KVLoraRank}, []int32{B, L, cfg.KVLoraRank + cfg.QKRopeHeadDim})
+
+	kPE = mlx.Reshape(kPE, B, L, 1, cfg.QKRopeHeadDim)
+	kPE = mlx.Transpose(kPE, 0, 2, 1, 3)
+
+	kvLatent := a.KVALayerNorm.Forward(kvCompressed, cfg.RMSNormEps)
+	kvLatent = mlx.ExpandDims(kvLatent, 1)
+
+	qPE = mlx.RoPEWithBase(qPE, int(cfg.QKRopeHeadDim), true, cfg.RopeTheta, 1.0, positions)
+	kPE = mlx.RoPEWithBase(kPE, int(cfg.QKRopeHeadDim), true, cfg.RopeTheta, 1.0, positions)
+
+	qLatent := a.EmbedQ.Forward(qNope)
+
+	keys := mlx.Concatenate([]*mlx.Array{kvLatent, kPE}, 3)
+
+	// MLA compresses K and V into a single tensor: the cache stores
+	// [kvLatent, kPE] concatenated along the last dim as its keys,
+	// and V is the kvLatent prefix (first KVLoraRank positions) of
+	// that same tensor. WithMLAHistory handles the slice on our
+	// behalf so the model never touches the history's K/V.
+	var kv nn.SDPAOption
+	if c != nil {
+		placeholderValues := mlx.ZerosF32([]int32{B, 1, L, 0})
+		history := c.(cache.Attention).Update(b, keys, placeholderValues)
+		kv = nn.WithMLAHistory(history, int(cfg.KVLoraRank))
+	} else {
+		values := mlx.SliceStartStop(keys, []int32{0, 0, 0, 0}, []int32{B, 1, L, cfg.KVLoraRank})
+		kv = nn.WithKV(keys, values, b.SeqQueryLens)
+	}
+
+	queries := mlx.Concatenate([]*mlx.Array{qLatent, qPE}, 3)
+
+	out := nn.ScaledDotProductAttention(b, queries, cfg.Scale, kv, nn.WithMask(nn.CausalMask()))
+	out = a.UnembedOut.Forward(out)
+
+	out = mlx.Reshape(mlx.Transpose(out, 0, 2, 1, 3), B, L, cfg.NumAttentionHeads*cfg.VHeadDim)
+
+	return a.OProj.Forward(out)
+}
+
+// DenseMLP implements the standard SwiGLU MLP for dense layers
+type DenseMLP struct {
+	GateProj nn.LinearLayer
+	UpProj   nn.LinearLayer
+	DownProj nn.LinearLayer
+}
+
+// Forward applies the SwiGLU MLP
+func (m *DenseMLP) Forward(x *mlx.Array) *mlx.Array {
+	return m.DownProj.Forward(mlx.SwiGLU(m.GateProj.Forward(x), m.UpProj.Forward(x)))
+}
+
+// MoEGate implements the expert gating mechanism
+type MoEGate struct {
+	Gate                 nn.LinearLayer
+	EScoreCorrectionBias *mlx.Array
+}
+
+// Forward computes expert selection indices and scores
+func (g *MoEGate) Forward(x *mlx.Array, cfg *Config) (*mlx.Array, *mlx.Array) {
+	gates := g.Gate.Forward(x)
+
+	var origScores, negScores *mlx.Array
+	if g.EScoreCorrectionBias != nil {
+		origScores, negScores = mlx.SigmoidRouter(gates, g.EScoreCorrectionBias)
+	} else {
+		origScores = mlx.Sigmoid(gates)
+		negScores = mlx.Neg(origScores)
+	}
+
+	topK := cfg.NumExpertsPerTok
+	inds := mlx.Argpartition(negScores, int(topK)-1, -1)
+
+	dims := inds.Dims()
+	inds = mlx.SliceStartStop(inds, []int32{0, 0, 0}, []int32{int32(dims[0]), int32(dims[1]), topK})
+
+	scores := mlx.TakeAlongAxis(origScores, inds, -1)
+
+	if topK > 1 && cfg.NormTopKProb {
+		sumScores := mlx.Sum(scores, -1, true)
+		scores = mlx.Div(scores, sumScores)
+	}
+
+	scores = mlx.MulScalar(scores, cfg.RoutedScalingFactor)
+
+	return inds, scores
+}
+
+// SwitchMLP implements the MoE expert computation using stacked weights
+type SwitchMLP struct {
+	GateWeight *mlx.Array
+	UpWeight   *mlx.Array
+	DownWeight *mlx.Array
+
+	GateWeightQ, GateScales, GateBiases *mlx.Array
+	UpWeightQ, UpScales, UpBiases       *mlx.Array
+	DownWeightQ, DownScales, DownBiases *mlx.Array
+
+	GateBits int
+	UpBits   int
+	DownBits int
+
+	GateGroupSize int
+	UpGroupSize   int
+	DownGroupSize int
+
+	UseQuantized bool
+}
+
+// Forward applies the switched expert MLP
+func (s *SwitchMLP) Forward(x *mlx.Array, indices *mlx.Array, cfg *Config) *mlx.Array {
+	dims := x.Dims()
+	B, L := int32(dims[0]), int32(dims[1])
+	topK := cfg.NumExpertsPerTok
+
+	xExpanded := mlx.ExpandDims(mlx.ExpandDims(x, -2), -2)
+
+	xFlat := mlx.Reshape(xExpanded, B*L, 1, 1, cfg.HiddenSize)
+
+	idxFlat := mlx.Reshape(indices, B*L, topK)
+
+	doSort := B*L >= 64
+	var invOrder *mlx.Array
+	n := B * L * topK
+
+	if doSort {
+		idxAll := mlx.Flatten(idxFlat)
+		order := mlx.Argsort(idxAll, 0)
+		invOrder = mlx.Argsort(order, 0)
+		xFlat = mlx.ExpandDims(mlx.Take(mlx.Squeeze(xFlat, 1), mlx.FloorDivideScalar(order, topK), 0), 1)
+		idxFlat = mlx.Reshape(mlx.Take(idxAll, order, 0), n, 1)
+	}
+
+	var gate, up, hidden, down *mlx.Array
+
+	if s.UseQuantized {
+		gate = mlx.GatherQMM(xFlat, s.GateWeightQ, s.GateScales, s.GateBiases,
+			nil, idxFlat, true, s.GateGroupSize, s.GateBits, cfg.QuantMode, doSort)
+		up = mlx.GatherQMM(xFlat, s.UpWeightQ, s.UpScales, s.UpBiases,
+			nil, idxFlat, true, s.UpGroupSize, s.UpBits, cfg.QuantMode, doSort)
+
+		hidden = mlx.SwiGLU(gate, up)
+
+		down = mlx.GatherQMM(hidden, s.DownWeightQ, s.DownScales, s.DownBiases,
+			nil, idxFlat, true, s.DownGroupSize, s.DownBits, cfg.QuantMode, doSort)
+	} else {
+		gate = mlx.GatherMM(xFlat, mlx.Transpose(s.GateWeight, 0, 2, 1), nil, idxFlat, doSort)
+		up = mlx.GatherMM(xFlat, mlx.Transpose(s.UpWeight, 0, 2, 1), nil, idxFlat, doSort)
+
+		hidden = mlx.SwiGLU(gate, up)
+
+		down = mlx.GatherMM(hidden, mlx.Transpose(s.DownWeight, 0, 2, 1), nil, idxFlat, doSort)
+	}
+
+	if doSort {
+		down = mlx.Reshape(mlx.Take(mlx.Squeeze(mlx.Squeeze(down, 2), 1), invOrder, 0), B*L, topK, cfg.HiddenSize)
+	} else {
+		down = mlx.Squeeze(down, 2)
+	}
+
+	return mlx.Reshape(down, B, L, topK, cfg.HiddenSize)
+}
+
+// SharedExperts implements the shared expert MLP
+type SharedExperts struct {
+	GateProj nn.LinearLayer
+	UpProj   nn.LinearLayer
+	DownProj nn.LinearLayer
+}
+
+// Forward applies the shared expert MLP
+func (s *SharedExperts) Forward(x *mlx.Array) *mlx.Array {
+	return s.DownProj.Forward(mlx.SwiGLU(s.GateProj.Forward(x), s.UpProj.Forward(x)))
+}
+
+// MoE implements the full Mixture of Experts layer
+type MoE struct {
+	Gate          *MoEGate
+	SwitchMLP     *SwitchMLP
+	SharedExperts *SharedExperts
+}
+
+// Forward applies the MoE layer
+func (m *MoE) Forward(x *mlx.Array, cfg *Config) *mlx.Array {
+	dims := x.Dims()
+	B, L := int32(dims[0]), int32(dims[1])
+
+	inds, scores := m.Gate.Forward(x, cfg)
+
+	expertOut := m.SwitchMLP.Forward(x, inds, cfg)
+
+	scoresExpanded := mlx.ExpandDims(scores, -1)
+	y := mlx.Sum(mlx.Mul(expertOut, scoresExpanded), 2, false)
+
+	if m.SharedExperts != nil {
+		y = mlx.Add(y, m.SharedExperts.Forward(x))
+	}
+
+	return mlx.Reshape(y, B, L, cfg.HiddenSize)
+}
+
+// DenseBlock represents a dense transformer block (for first_k_dense_replace layers)
+type DenseBlock struct {
+	Attention              *MLAAttention
+	MLP                    *DenseMLP
+	InputLayerNorm         *nn.RMSNorm
+	PostAttentionLayerNorm *nn.RMSNorm
+}
+
+// Forward applies the dense block
+func (blk *DenseBlock) Forward(x *mlx.Array, b *batch.Batch, c cache.Cache, positions *mlx.Array, B, L int32, cfg *Config) *mlx.Array {
+	r := blk.Attention.Forward(blk.InputLayerNorm.Forward(x, cfg.RMSNormEps), b, c, positions, B, L, cfg)
+	h := mlx.Add(x, r)
+
+	r = blk.MLP.Forward(blk.PostAttentionLayerNorm.Forward(h, cfg.RMSNormEps))
+	return mlx.Add(h, r)
+}
+
+// MoEBlock represents a MoE transformer block
+type MoEBlock struct {
+	Attention              *MLAAttention
+	MoE                    *MoE
+	InputLayerNorm         *nn.RMSNorm
+	PostAttentionLayerNorm *nn.RMSNorm
+}
+
+// Forward applies the MoE block
+func (blk *MoEBlock) Forward(x *mlx.Array, b *batch.Batch, c cache.Cache, positions *mlx.Array, B, L int32, cfg *Config) *mlx.Array {
+	r := blk.Attention.Forward(blk.InputLayerNorm.Forward(x, cfg.RMSNormEps), b, c, positions, B, L, cfg)
+	h := mlx.Add(x, r)
+
+	r = blk.MoE.Forward(blk.PostAttentionLayerNorm.Forward(h, cfg.RMSNormEps), cfg)
+	return mlx.Add(h, r)
+}
+
+// Block interface for both dense and MoE blocks
+type Block interface {
+	Forward(x *mlx.Array, b *batch.Batch, c cache.Cache, positions *mlx.Array, B, L int32, cfg *Config) *mlx.Array
+}
+
+// Model represents the complete GLM4-MoE-Lite model
+type Model struct {
+	EmbedTokens nn.EmbeddingLayer
+	Layers      []Block
+	Norm        *nn.RMSNorm
+	LMHead      nn.LinearLayer
+
+	tok *tokenizer.Tokenizer
+	*Config
+}
+
+// computeScale computes the attention scale.
+func computeScale(cfg *Config) float32 {
+	keyLength := cfg.QKNopeHeadDim + cfg.QKRopeHeadDim
+	scale := float32(1.0 / math.Sqrt(float64(keyLength)))
+	if cfg.RopeScaling != nil && cfg.RopeScaling.MscaleAllDim > 0 && cfg.RopeScaling.Factor > 1 {
+		s := 0.1*cfg.RopeScaling.MscaleAllDim*float32(math.Log(float64(cfg.RopeScaling.Factor))) + 1.0
+		scale *= s * s
+	}
+	return scale
+}
+
+// supportsGatherQMM returns true if the quantization mode has GatherQMM kernel support.
+func supportsGatherQMM(mode string, bits int) bool {
+	return mode == "affine" && (bits == 4 || bits == 8)
+}
+
+// ExpertWeight holds a single expert's weight with optional quantization components.
+type ExpertWeight struct {
+	Weight    *mlx.Array
+	Scales    *mlx.Array
+	Biases    *mlx.Array
+	Bits      int
+	GroupSize int
+}
+
+// loadExpertWeight loads an expert weight from the tensor map.
+func loadExpertWeight(tensors map[string]*mlx.Array, path string, useQuantized bool, cfg *Config) *ExpertWeight {
+	w := tensors[path+".weight"]
+	if w == nil {
+		return nil
+	}
+
+	scales := tensors[path+".weight_scale"]
+	if scales != nil {
+		qbiases := tensors[path+".weight_qbias"]
+
+		groupSize, bits, mode := model.ResolveLinearQuantParams(
+			cfg.QuantGroupSize,
+			cfg.QuantBits,
+			cfg.QuantMode,
+			cfg.TensorQuant,
+			path+".weight",
+			w,
+			scales,
+		)
+
+		if useQuantized && supportsGatherQMM(mode, bits) {
+			return &ExpertWeight{Weight: w, Scales: scales, Biases: qbiases, Bits: bits, GroupSize: groupSize}
+		}
+
+		return &ExpertWeight{Weight: mlx.Dequantize(w, scales, qbiases, groupSize, bits, mode)}
+	}
+
+	return &ExpertWeight{Weight: w}
+}
+
+// StackedExpertWeights holds stacked weights for all experts.
+type StackedExpertWeights struct {
+	Weight    *mlx.Array
+	Scales    *mlx.Array
+	Biases    *mlx.Array
+	Bits      int
+	GroupSize int
+}
+
+// collectAndStackExpertWeights loads and stacks expert weights for one projection type.
+func collectAndStackExpertWeights(
+	tensors map[string]*mlx.Array,
+	prefix string,
+	projName string,
+	numExperts int32,
+	useQuantized bool,
+	cfg *Config,
+) *StackedExpertWeights {
+	var w, s, b []*mlx.Array
+	var bits, groupSize int
+
+	for e := int32(0); e < numExperts; e++ {
+		path := fmt.Sprintf("%s.mlp.experts.%d.%s", prefix, e, projName)
+		ew := loadExpertWeight(tensors, path, useQuantized, cfg)
+		if ew == nil {
+			continue
+		}
+		w = append(w, ew.Weight)
+		if ew.Scales != nil {
+			s = append(s, ew.Scales)
+		}
+		if ew.Biases != nil {
+			b = append(b, ew.Biases)
+		}
+		if e == 0 {
+			bits = ew.Bits
+			groupSize = ew.GroupSize
+		}
+	}
+
+	result := &StackedExpertWeights{Bits: bits, GroupSize: groupSize}
+	if len(w) > 0 {
+		result.Weight = mlx.Stack(w, 0)
+		if len(s) > 0 {
+			result.Scales = mlx.Stack(s, 0)
+		}
+		if len(b) > 0 {
+			result.Biases = mlx.Stack(b, 0)
+		}
+	}
+	return result
+}
+
+// sanitizeExpertWeights stacks individual expert weights into tensors.
+func sanitizeExpertWeights(tensors map[string]*mlx.Array, prefix string, numExperts int32, useQuantized bool, cfg *Config) (gate, up, down *StackedExpertWeights) {
+	gate = collectAndStackExpertWeights(tensors, prefix, "gate_proj", numExperts, useQuantized, cfg)
+	up = collectAndStackExpertWeights(tensors, prefix, "up_proj", numExperts, useQuantized, cfg)
+	down = collectAndStackExpertWeights(tensors, prefix, "down_proj", numExperts, useQuantized, cfg)
+	return gate, up, down
+}
+
+// sanitizeMLAWeights transforms kv_b_proj weights into absorbed MLA format.
+func sanitizeMLAWeights(tensors map[string]*mlx.Array, prefix string, cfg *Config) (*mlx.Array, *mlx.Array) {
+	path := prefix + ".self_attn.kv_b_proj"
+	w := tensors[path+".weight"]
+	if w == nil {
+		return nil, nil
+	}
+
+	// Check if quantized and dequantize
+	if scales := tensors[path+".weight_scale"]; scales != nil {
+		qbiases := tensors[path+".weight_qbias"]
+		groupSize, bits, mode := model.ResolveLinearQuantParams(
+			cfg.QuantGroupSize,
+			cfg.QuantBits,
+			cfg.QuantMode,
+			cfg.TensorQuant,
+			path+".weight",
+			w,
+			scales,
+		)
+		w = mlx.Dequantize(w, scales, qbiases, groupSize, bits, mode)
+	}
+
+	headDim := cfg.QKNopeHeadDim + cfg.VHeadDim
+	w = mlx.Reshape(w, cfg.NumAttentionHeads, headDim, cfg.KVLoraRank)
+
+	wk := mlx.SliceStartStop(w, []int32{0, 0, 0}, []int32{cfg.NumAttentionHeads, cfg.QKNopeHeadDim, cfg.KVLoraRank})
+	wv := mlx.SliceStartStop(w, []int32{0, cfg.QKNopeHeadDim, 0}, []int32{cfg.NumAttentionHeads, headDim, cfg.KVLoraRank})
+
+	embedQ := mlx.Transpose(wk, 0, 2, 1)
+	unembedOut := wv
+
+	return embedQ, unembedOut
+}
+
+// newModel creates a new GLM4-MoE-Lite model from a Root (config + tokenizer,
+// no weights loaded yet). Called by the registry via base.New().
+func newModel(root *model.Root) (base.Model, error) {
+	configData, err := root.Manifest.ReadConfig("config.json")
+	if err != nil {
+		return nil, fmt.Errorf("load config: %w", err)
+	}
+
+	var cfg Config
+	if err := json.Unmarshal(configData, &cfg); err != nil {
+		return nil, fmt.Errorf("parse config: %w", err)
+	}
+
+	cfg.QHeadDim = cfg.QKNopeHeadDim + cfg.QKRopeHeadDim
+	cfg.Scale = computeScale(&cfg)
+
+	// Set up quantization parameters from pre-scanned metadata
+	if qt := root.QuantType(); qt != "" {
+		cfg.QuantGroupSize, cfg.QuantBits, cfg.QuantMode = model.QuantizationParams(qt)
+		if gs := root.GroupSize(); gs > 0 {
+			cfg.QuantGroupSize = gs
+		}
+	} else {
+		cfg.QuantGroupSize, cfg.QuantBits, cfg.QuantMode = model.QuantizationParams("")
+	}
+	cfg.TensorQuant = root.AllTensorQuant()
+
+	// Load tokenizer
+	tokData, err := root.Manifest.ReadConfig("tokenizer.json")
+	if err != nil {
+		return nil, fmt.Errorf("load tokenizer config: %w", err)
+	}
+
+	tokConfig := &tokenizer.TokenizerConfig{
+		ConfigJSON: configData,
+	}
+
+	if genConfigData, err := root.Manifest.ReadConfig("generation_config.json"); err == nil {
+		tokConfig.GenerationConfigJSON = genConfigData
+	}
+
+	if tokConfigData, err := root.Manifest.ReadConfig("tokenizer_config.json"); err == nil {
+		tokConfig.TokenizerConfigJSON = tokConfigData
+	}
+
+	tok, err := tokenizer.LoadFromBytesWithConfig(tokData, tokConfig)
+	if err != nil {
+		return nil, fmt.Errorf("parse tokenizer: %w", err)
+	}
+
+	m := &Model{
+		Layers: make([]Block, cfg.NumHiddenLayers),
+		Config: &cfg,
+		tok:    tok,
+	}
+
+	return m, nil
+}
+
+// LoadWeights receives all tensors loaded from the manifest and assigns them
+// to model fields. Handles MLA absorption, expert stacking, and quantized
+// layer creation.
+func (m *Model) LoadWeights(tensors map[string]*mlx.Array) error {
+	cfg := m.Config
+	linears := model.NewLinearFactory(tensors, cfg.QuantGroupSize, cfg.QuantBits, cfg.QuantMode, cfg.TensorQuant)
+	useQuantized := supportsGatherQMM(cfg.QuantMode, cfg.QuantBits)
+	if !useQuantized && cfg.TensorQuant != nil {
+		for _, tq := range cfg.TensorQuant {
+			if tq == nil {
+				continue
+			}
+			_, bits, mode := model.QuantizationParams(tq.QuantType)
+			if supportsGatherQMM(mode, bits) {
+				useQuantized = true
+				break
+			}
+		}
+	}
+
+	// Load embedding
+	m.EmbedTokens = model.MakeEmbeddingLayer(tensors, "model.embed_tokens", cfg.QuantGroupSize, cfg.QuantBits, cfg.QuantMode, cfg.TensorQuant)
+
+	// Load final norm
+	if w := tensors["model.norm.weight"]; w != nil {
+		m.Norm = nn.NewRMSNorm(w, cfg.RMSNormEps)
+	}
+
+	// Load LM head
+	m.LMHead = linears.Make("lm_head")
+
+	// Load layers
+	for i := int32(0); i < cfg.NumHiddenLayers; i++ {
+		prefix := fmt.Sprintf("model.layers.%d", i)
+
+		// Load attention (same for both block types)
+		attn := &MLAAttention{}
+		attn.QAProj = linears.Make(prefix + ".self_attn.q_a_proj")
+		if w := tensors[prefix+".self_attn.q_a_layernorm.weight"]; w != nil {
+			attn.QALayerNorm = nn.NewRMSNorm(w, cfg.RMSNormEps)
+		}
+		attn.QBProj = linears.Make(prefix + ".self_attn.q_b_proj")
+		attn.KVAProjWithMQA = linears.Make(prefix + ".self_attn.kv_a_proj_with_mqa")
+		if w := tensors[prefix+".self_attn.kv_a_layernorm.weight"]; w != nil {
+			attn.KVALayerNorm = nn.NewRMSNorm(w, cfg.RMSNormEps)
+		}
+		attn.OProj = linears.Make(prefix + ".self_attn.o_proj")
+
+		// Sanitize MLA weights for absorbed attention
+		embedQ, unembedOut := sanitizeMLAWeights(tensors, prefix, cfg)
+		attn.EmbedQ = nn.NewMultiLinear(embedQ)
+		attn.UnembedOut = nn.NewMultiLinear(unembedOut)
+
+		inputLN := tensors[prefix+".input_layernorm.weight"]
+		postAttnLN := tensors[prefix+".post_attention_layernorm.weight"]
+
+		if i < cfg.FirstKDenseReplace {
+			// Dense block
+			block := &DenseBlock{Attention: attn}
+			if inputLN != nil {
+				block.InputLayerNorm = nn.NewRMSNorm(inputLN, cfg.RMSNormEps)
+			}
+			if postAttnLN != nil {
+				block.PostAttentionLayerNorm = nn.NewRMSNorm(postAttnLN, cfg.RMSNormEps)
+			}
+
+			block.MLP = &DenseMLP{
+				GateProj: linears.Make(prefix + ".mlp.gate_proj"),
+				UpProj:   linears.Make(prefix + ".mlp.up_proj"),
+				DownProj: linears.Make(prefix + ".mlp.down_proj"),
+			}
+
+			m.Layers[i] = block
+		} else {
+			// MoE block
+			block := &MoEBlock{Attention: attn}
+			if inputLN != nil {
+				block.InputLayerNorm = nn.NewRMSNorm(inputLN, cfg.RMSNormEps)
+			}
+			if postAttnLN != nil {
+				block.PostAttentionLayerNorm = nn.NewRMSNorm(postAttnLN, cfg.RMSNormEps)
+			}
+
+			// Stack expert weights
+			gate, up, down := sanitizeExpertWeights(tensors, prefix, cfg.NRoutedExperts, useQuantized, cfg)
+
+			switchMLP := &SwitchMLP{UseQuantized: useQuantized}
+			if useQuantized {
+				switchMLP.GateWeightQ = gate.Weight
+				switchMLP.GateScales = gate.Scales
+				switchMLP.GateBiases = gate.Biases
+				switchMLP.GateBits = gate.Bits
+				switchMLP.GateGroupSize = gate.GroupSize
+				switchMLP.UpWeightQ = up.Weight
+				switchMLP.UpScales = up.Scales
+				switchMLP.UpBiases = up.Biases
+				switchMLP.UpBits = up.Bits
+				switchMLP.UpGroupSize = up.GroupSize
+				switchMLP.DownWeightQ = down.Weight
+				switchMLP.DownScales = down.Scales
+				switchMLP.DownBiases = down.Biases
+				switchMLP.DownBits = down.Bits
+				switchMLP.DownGroupSize = down.GroupSize
+			} else {
+				switchMLP.GateWeight = gate.Weight
+				switchMLP.UpWeight = up.Weight
+				switchMLP.DownWeight = down.Weight
+			}
+
+			moeGate := &MoEGate{}
+			moeGate.Gate = linears.Make(prefix + ".mlp.gate")
+			if bias := tensors[prefix+".mlp.gate.e_score_correction_bias"]; bias != nil {
+				moeGate.EScoreCorrectionBias = bias
+			}
+
+			block.MoE = &MoE{
+				Gate:      moeGate,
+				SwitchMLP: switchMLP,
+			}
+
+			// Load shared experts if present
+			if cfg.NSharedExperts > 0 {
+				block.MoE.SharedExperts = &SharedExperts{
+					GateProj: linears.Make(prefix + ".mlp.shared_experts.gate_proj"),
+					UpProj:   linears.Make(prefix + ".mlp.shared_experts.up_proj"),
+					DownProj: linears.Make(prefix + ".mlp.shared_experts.down_proj"),
+				}
+			}
+
+			m.Layers[i] = block
+		}
+	}
+
+	return nil
+}
+
+// Forward computes the forward pass of the model
+func (m *Model) Forward(b *batch.Batch, caches []cache.Cache) *mlx.Array {
+	dims := b.InputIDs.Dims()
+	B, L := int32(dims[0]), int32(dims[1])
+	positions := mlx.FromValues(b.SeqOffsets, len(b.SeqOffsets))
+
+	h := m.EmbedTokens.Forward(b.InputIDs)
+
+	for i, layer := range m.Layers {
+		var c cache.Cache
+		if caches != nil {
+			c = caches[i]
+		}
+		h = layer.Forward(h, b, c, positions, B, L, m.Config)
+	}
+
+	h = m.Norm.Forward(h, m.RMSNormEps)
+	return h
+}
+
+// Unembed applies the LM head to get logits.
+func (m *Model) Unembed(x *mlx.Array) *mlx.Array {
+	return m.LMHead.Forward(x)
+}
+
+// NumLayers returns the number of transformer layers
+func (m *Model) NumLayers() int { return len(m.Layers) }
+
+// MaxContextLength returns the maximum context length
+func (m *Model) MaxContextLength() int { return int(m.MaxPositionEmbeddings) }
+
+// VocabSize returns the vocabulary size
+func (m *Model) VocabSize() int32 { return m.Config.VocabSize }
+
+// Tokenizer returns the model's tokenizer
+func (m *Model) Tokenizer() *tokenizer.Tokenizer { return m.tok }
+
+// NewCache creates a new KV cache for the model
+func (m *Model) NewCache(maxSeqLen int32) []cache.Cache {
+	caches := make([]cache.Cache, len(m.Layers))
+	for i := range caches {
+		caches[i] = cache.NewKVCache()
+	}
+	return caches
+}
+
+// FormatPrompt applies the GLM-4 chat template with thinking enabled by default.
+func (m *Model) FormatPrompt(prompt string) string {
+	return "[gMASK]<sop><|user|>" + prompt + "<|assistant|><think>"
+}
+
+// FormatPromptWithThinking applies the GLM-4 chat template with explicit thinking control.
+func (m *Model) FormatPromptWithThinking(prompt string, think bool) string {
+	if think {
+		return "[gMASK]<sop><|user|>" + prompt + "<|assistant|><think>"
+	}
+	return "[gMASK]<sop><|user|>" + prompt + "<|assistant|></think>"
+}
+
+// NewRenderer returns a new Renderer for formatting multi-turn conversations.
+func (m *Model) NewRenderer() *Renderer {
+	return &Renderer{}
+}
+
+// NewParser returns a new Parser for extracting thinking and tool calls from output.
+func (m *Model) NewParser() *Parser {
+	return &Parser{}
+}
--- a/x/models/glm4_moe_lite/parser.go
+++ b/x/models/glm4_moe_lite/parser.go
@@ -0,0 +1,520 @@
+package glm4_moe_lite
+
+import (
+	"context"
+	"encoding/json"
+	"encoding/xml"
+	"fmt"
+	"log/slog"
+	"strings"
+	"unicode"
+
+	"github.com/ollama/ollama/api"
+	"github.com/ollama/ollama/logutil"
+)
+
+type parserState int
+
+const (
+	parserState_LookingForThinkingOpen parserState = iota
+	parserState_ThinkingStartedEatingWhitespace
+	parserState_CollectingThinking
+	parserState_ThinkingDoneEatingWhitespace
+	parserState_CollectingContent
+	parserState_ToolStartedEatingWhitespace
+	parserState_CollectingToolContent
+)
+
+const (
+	thinkingOpenTag  = "<think>"
+	thinkingCloseTag = "</think>"
+	toolOpenTag      = "<tool_call>"
+	toolCloseTag     = "</tool_call>"
+)
+
+// Parser parses GLM4-MoE-Lite model output to extract thinking and tool calls.
+// GLM-4's prompt ends with <think> when thinking is enabled, so the parser
+// must start in CollectingThinking state (the model outputs thinking content directly).
+type Parser struct {
+	state  parserState
+	buffer strings.Builder
+	tools  []api.Tool
+}
+
+// HasToolSupport returns true as GLM4 supports tool calling.
+func (p *Parser) HasToolSupport() bool {
+	return true
+}
+
+// HasThinkingSupport returns true as GLM4 supports thinking mode.
+func (p *Parser) HasThinkingSupport() bool {
+	return true
+}
+
+// Init initializes the parser with tools and thinking configuration.
+func (p *Parser) Init(tools []api.Tool, lastMessage *api.Message, thinkValue *api.ThinkValue) []api.Tool {
+	p.tools = tools
+	// When thinking is enabled (nil or true), the prompt ends with <think>,
+	// so model output starts directly with thinking content (no opening tag).
+	if thinkValue == nil || thinkValue.Bool() {
+		p.state = parserState_CollectingThinking
+	}
+	return tools
+}
+
+type parserEvent interface {
+	isParserEvent()
+}
+
+type eventContent struct {
+	content string
+}
+
+func (eventContent) isParserEvent() {}
+
+type eventRawToolCall struct {
+	raw string
+}
+
+func (eventRawToolCall) isParserEvent() {}
+
+type eventThinkingContent struct {
+	content string
+}
+
+func (eventThinkingContent) isParserEvent() {}
+
+// Add processes new output text and returns parsed content, thinking, and tool calls.
+func (p *Parser) Add(s string, done bool) (content string, thinking string, calls []api.ToolCall, err error) {
+	p.buffer.WriteString(s)
+	events := p.parseEvents()
+
+	var toolCalls []api.ToolCall
+	var contentSb strings.Builder
+	var thinkingSb strings.Builder
+
+	for _, event := range events {
+		switch event := event.(type) {
+		case eventRawToolCall:
+			toolCall, err := parseToolCall(event, p.tools)
+			if err != nil {
+				slog.Warn("glm-4 tool call parsing failed", "error", err)
+				return "", "", nil, err
+			}
+			toolCalls = append(toolCalls, toolCall)
+		case eventThinkingContent:
+			thinkingSb.WriteString(event.content)
+		case eventContent:
+			contentSb.WriteString(event.content)
+		}
+	}
+
+	return contentSb.String(), thinkingSb.String(), toolCalls, nil
+}
+
+func (p *Parser) parseEvents() []parserEvent {
+	var all []parserEvent
+
+	keepLooping := true
+	for keepLooping {
+		var events []parserEvent
+		events, keepLooping = p.eat()
+		if len(events) > 0 {
+			all = append(all, events...)
+		}
+	}
+
+	if len(all) > 0 {
+		slog.Log(context.TODO(), logutil.LevelTrace, "glm-4 events parsed", "events", all, "state", p.state, "buffer", p.buffer.String())
+	}
+
+	return all
+}
+
+// eatLeadingWhitespaceAndTransitionTo consumes leading whitespace from the buffer
+// and transitions to the next state. Returns (nil, false) if only whitespace remains
+// in the buffer (needs more input), or (nil, true) if we successfully transitioned.
+func (p *Parser) eatLeadingWhitespaceAndTransitionTo(nextState parserState) ([]parserEvent, bool) {
+	trimmed := strings.TrimLeftFunc(p.buffer.String(), unicode.IsSpace)
+	p.buffer.Reset()
+	if trimmed == "" {
+		return nil, false // Still only whitespace, keep waiting for more input
+	}
+	p.state = nextState
+	p.buffer.WriteString(trimmed)
+	return nil, true // Successfully transitioned
+}
+
+// splitAtTag splits the buffer at the given tag, returns the content before (trimmed of trailing whitespace),
+// the content after (optionally trimmed of leading whitespace), and updates the buffer
+func (p *Parser) splitAtTag(tag string, trimAfter bool) (string, string) {
+	split := strings.SplitN(p.buffer.String(), tag, 2)
+	before := split[0]
+	before = strings.TrimRightFunc(before, unicode.IsSpace)
+	after := split[1]
+	if trimAfter {
+		after = strings.TrimLeftFunc(after, unicode.IsSpace)
+	}
+	p.buffer.Reset()
+	p.buffer.WriteString(after)
+	return before, after
+}
+
+func (p *Parser) eat() ([]parserEvent, bool) {
+	var events []parserEvent
+
+	switch p.state {
+	case parserState_LookingForThinkingOpen:
+		trimmed := strings.TrimLeftFunc(p.buffer.String(), unicode.IsSpace)
+		if strings.HasPrefix(trimmed, thinkingOpenTag) {
+			// Found <think> opening tag
+			after := strings.TrimPrefix(trimmed, thinkingOpenTag)
+			after = strings.TrimLeftFunc(after, unicode.IsSpace)
+			p.buffer.Reset()
+			p.buffer.WriteString(after)
+			if after == "" {
+				p.state = parserState_ThinkingStartedEatingWhitespace
+			} else {
+				p.state = parserState_CollectingThinking
+			}
+			return events, true
+		} else if strings.HasPrefix(thinkingOpenTag, trimmed) {
+			// Partial opening tag seen, keep accumulating
+			return events, false
+		} else if trimmed == "" {
+			// Only whitespace, keep accumulating
+			return events, false
+		} else {
+			// No thinking tag found, skip to content collection
+			p.state = parserState_CollectingContent
+			// Don't trim - we want to keep the original content
+			return events, true
+		}
+
+	case parserState_ThinkingStartedEatingWhitespace:
+		return p.eatLeadingWhitespaceAndTransitionTo(parserState_CollectingThinking)
+
+	case parserState_CollectingThinking:
+		acc := p.buffer.String()
+		if strings.Contains(acc, thinkingCloseTag) {
+			thinking, remaining := p.splitAtTag(thinkingCloseTag, true)
+			if len(thinking) > 0 {
+				events = append(events, eventThinkingContent{content: thinking})
+			}
+			if remaining == "" {
+				p.state = parserState_ThinkingDoneEatingWhitespace
+			} else {
+				p.state = parserState_CollectingContent
+			}
+			return events, true
+		} else if overlapLen := overlap(acc, thinkingCloseTag); overlapLen > 0 {
+			// Partial closing tag - withhold it along with any trailing whitespace before it
+			beforePartialTag := acc[:len(acc)-overlapLen]
+			trailingWsLen := trailingWhitespaceLen(beforePartialTag)
+			ambiguousStart := len(beforePartialTag) - trailingWsLen
+
+			unambiguous := acc[:ambiguousStart]
+			ambiguous := acc[ambiguousStart:]
+			p.buffer.Reset()
+			p.buffer.WriteString(ambiguous)
+			if len(unambiguous) > 0 {
+				events = append(events, eventThinkingContent{content: unambiguous})
+			}
+			return events, false
+		} else {
+			// Pure thinking content - withhold trailing whitespace (might precede closing tag)
+			whitespaceLen := trailingWhitespaceLen(acc)
+			ambiguousStart := len(acc) - whitespaceLen
+
+			unambiguous := acc[:ambiguousStart]
+			ambiguous := acc[ambiguousStart:]
+			p.buffer.Reset()
+			p.buffer.WriteString(ambiguous)
+			if len(unambiguous) > 0 {
+				events = append(events, eventThinkingContent{content: unambiguous})
+			}
+			return events, false
+		}
+
+	case parserState_ThinkingDoneEatingWhitespace:
+		return p.eatLeadingWhitespaceAndTransitionTo(parserState_CollectingContent)
+
+	case parserState_CollectingContent:
+		if strings.Contains(p.buffer.String(), toolOpenTag) {
+			before, after := p.splitAtTag(toolOpenTag, true)
+			if len(before) > 0 {
+				events = append(events, eventContent{content: before})
+			}
+			if after == "" {
+				p.state = parserState_ToolStartedEatingWhitespace
+			} else {
+				p.state = parserState_CollectingToolContent
+			}
+			return events, true
+		} else if overlapLen := overlap(p.buffer.String(), toolOpenTag); overlapLen > 0 {
+			beforePartialTag := p.buffer.String()[:len(p.buffer.String())-overlapLen]
+			trailingWsLen := trailingWhitespaceLen(beforePartialTag)
+			ambiguousStart := len(beforePartialTag) - trailingWsLen
+
+			unambiguous := p.buffer.String()[:ambiguousStart]
+			ambiguous := p.buffer.String()[ambiguousStart:]
+			p.buffer.Reset()
+			p.buffer.WriteString(ambiguous)
+			if len(unambiguous) > 0 {
+				events = append(events, eventContent{content: unambiguous})
+			}
+			return events, false
+		} else {
+			whitespaceLen := trailingWhitespaceLen(p.buffer.String())
+			ambiguousStart := len(p.buffer.String()) - whitespaceLen
+
+			unambiguous := p.buffer.String()[:ambiguousStart]
+			ambiguous := p.buffer.String()[ambiguousStart:]
+			p.buffer.Reset()
+			p.buffer.WriteString(ambiguous)
+			if len(unambiguous) > 0 {
+				events = append(events, eventContent{content: unambiguous})
+			}
+			return events, false
+		}
+
+	case parserState_ToolStartedEatingWhitespace:
+		return p.eatLeadingWhitespaceAndTransitionTo(parserState_CollectingToolContent)
+
+	case parserState_CollectingToolContent:
+		acc := p.buffer.String()
+		if strings.Contains(acc, toolCloseTag) {
+			toolContent, _ := p.splitAtTag(toolCloseTag, true)
+			if len(toolContent) == 0 {
+				slog.Warn("glm4 tool call closing tag found but no content before it")
+			}
+			events = append(events, eventRawToolCall{raw: toolContent})
+			p.state = parserState_CollectingContent
+			return events, true
+		} else {
+			// Keep accumulating - tool calls are not streamed
+			// We just wait for the closing tag
+			return events, false
+		}
+
+	default:
+		panic("unreachable")
+	}
+}
+
+// overlap returns the length of the overlap between the end of s and the start of tag.
+func overlap(s, tag string) int {
+	for i := 1; i <= len(tag) && i <= len(s); i++ {
+		if strings.HasSuffix(s, tag[:i]) {
+			return i
+		}
+	}
+	return 0
+}
+
+// trailingWhitespaceLen returns the length of trailing whitespace in s.
+func trailingWhitespaceLen(s string) int {
+	trimmed := strings.TrimRightFunc(s, unicode.IsSpace)
+	return len(s) - len(trimmed)
+}
+
+// ToolCallXML represents the structure of a GLM-4 tool call for XML parsing
+type ToolCallXML struct {
+	XMLName xml.Name `xml:"tool_call"`
+	Content string   `xml:",chardata"` // Function name (text nodes between tags)
+	Keys    []string `xml:"arg_key"`   // All arg_key elements in document order
+	Values  []string `xml:"arg_value"` // All arg_value elements in document order
+}
+
+// escapeContent escapes XML entities in text content while preserving arg_key/arg_value tags
+func escapeContent(s string) string {
+	var result strings.Builder
+	inTag := false
+
+	for i := range len(s) {
+		ch := s[i]
+
+		if ch == '<' {
+			// Check if this is a known tag
+			if strings.HasPrefix(s[i:], "<arg_key>") ||
+				strings.HasPrefix(s[i:], "</arg_key>") ||
+				strings.HasPrefix(s[i:], "<arg_value>") ||
+				strings.HasPrefix(s[i:], "</arg_value>") {
+				inTag = true
+			}
+		}
+
+		if inTag {
+			result.WriteByte(ch)
+			if ch == '>' {
+				inTag = false
+			}
+		} else {
+			// Escape special characters in text content
+			switch ch {
+			case '&':
+				result.WriteString("&amp;")
+			case '<':
+				result.WriteString("&lt;")
+			case '>':
+				result.WriteString("&gt;")
+			default:
+				result.WriteByte(ch)
+			}
+		}
+	}
+
+	return result.String()
+}
+
+// repairUnclosedArgValues inserts missing </arg_value> closing tags.
+// GLM models sometimes omit the closing tag, producing XML like:
+//
+//	<arg_value>value</tool_call>
+//
+// instead of:
+//
+//	<arg_value>value</arg_value></tool_call>
+func repairUnclosedArgValues(s string) string {
+	var result strings.Builder
+	for {
+		openIdx := strings.Index(s, "<arg_value>")
+		if openIdx == -1 {
+			result.WriteString(s)
+			break
+		}
+		afterOpen := openIdx + len("<arg_value>")
+		closeIdx := strings.Index(s[afterOpen:], "</arg_value>")
+		nextKeyIdx := strings.Index(s[afterOpen:], "<arg_key>")
+		if closeIdx != -1 && (nextKeyIdx == -1 || closeIdx < nextKeyIdx) {
+			end := afterOpen + closeIdx + len("</arg_value>")
+			result.WriteString(s[:end])
+			s = s[end:]
+			continue
+		}
+		if nextKeyIdx != -1 {
+			insertAt := afterOpen + nextKeyIdx
+			result.WriteString(s[:insertAt])
+			result.WriteString("</arg_value>")
+			s = s[insertAt:]
+		} else {
+			result.WriteString(s)
+			result.WriteString("</arg_value>")
+			break
+		}
+	}
+	return result.String()
+}
+
+func parseToolCall(raw eventRawToolCall, tools []api.Tool) (api.ToolCall, error) {
+	// Escape any unescaped entities in text content
+	escaped := escapeContent(raw.raw)
+
+	// Wrap the content in a root element to make it valid XML
+	xmlString := "<tool_call>" + escaped + "</tool_call>"
+
+	// Parse XML into struct, retrying once with repaired XML if it fails
+	var parsed ToolCallXML
+	if err := xml.Unmarshal([]byte(xmlString), &parsed); err != nil {
+		parsed = ToolCallXML{}
+		repaired := "<tool_call>" + repairUnclosedArgValues(escaped) + "</tool_call>"
+		if err2 := xml.Unmarshal([]byte(repaired), &parsed); err2 != nil {
+			return api.ToolCall{}, fmt.Errorf("failed to parse XML: %w", err)
+		}
+	}
+
+	// Extract and trim function name
+	functionName := strings.TrimSpace(parsed.Content)
+	if functionName == "" {
+		return api.ToolCall{}, fmt.Errorf("empty function name")
+	}
+
+	// Verify keys and values are paired correctly
+	if len(parsed.Keys) != len(parsed.Values) {
+		return api.ToolCall{}, fmt.Errorf("mismatched arg_key and arg_value counts: %d keys, %d values", len(parsed.Keys), len(parsed.Values))
+	}
+
+	// Find the matching tool to get parameter types
+	var matchedTool *api.Tool
+	for i := range tools {
+		if tools[i].Function.Name == functionName {
+			matchedTool = &tools[i]
+			break
+		}
+	}
+
+	// Build arguments map by pairing keys and values
+	toolCall := api.ToolCall{
+		Function: api.ToolCallFunction{
+			Name:      functionName,
+			Arguments: api.NewToolCallFunctionArguments(),
+		},
+	}
+
+	for i := range parsed.Keys {
+		key := strings.TrimSpace(parsed.Keys[i])
+		value := parsed.Values[i] // Don't trim here - parseValue handles it
+
+		// Look up parameter type
+		var paramType api.PropertyType
+		if matchedTool != nil && matchedTool.Function.Parameters.Properties != nil {
+			if prop, ok := matchedTool.Function.Parameters.Properties.Get(key); ok {
+				// Handle anyOf by collecting all types from the union
+				if len(prop.AnyOf) > 0 {
+					for _, anyOfProp := range prop.AnyOf {
+						paramType = append(paramType, anyOfProp.Type...)
+					}
+				} else {
+					paramType = prop.Type
+				}
+			}
+		}
+
+		// Parse value with type coercion
+		toolCall.Function.Arguments.Set(key, parseValue(value, paramType))
+	}
+
+	return toolCall, nil
+}
+
+// parseValue parses a string value and coerces it to the appropriate type based on paramType.
+func parseValue(value string, paramType api.PropertyType) any {
+	value = strings.TrimSpace(value)
+
+	// If no type specified, return as string
+	if len(paramType) == 0 {
+		return value
+	}
+
+	// Try to parse based on specified types
+	for _, t := range paramType {
+		switch t {
+		case "boolean":
+			if value == "true" {
+				return true
+			}
+			if value == "false" {
+				return false
+			}
+		case "integer":
+			var i int64
+			if _, err := fmt.Sscanf(value, "%d", &i); err == nil {
+				return i
+			}
+		case "number":
+			var f float64
+			if _, err := fmt.Sscanf(value, "%f", &f); err == nil {
+				return f
+			}
+		case "array", "object":
+			// Try to parse as JSON
+			var result any
+			if err := json.Unmarshal([]byte(value), &result); err == nil {
+				return result
+			}
+		}
+	}
+
+	// Default to string
+	return value
+}
--- a/x/models/glm4_moe_lite/parser_test.go
+++ b/x/models/glm4_moe_lite/parser_test.go
@@ -0,0 +1,190 @@
+package glm4_moe_lite
+
+import (
+	"testing"
+
+	"github.com/ollama/ollama/api"
+)
+
+func TestParserThinking(t *testing.T) {
+	tests := []struct {
+		name          string
+		input         string
+		thinkEnabled  bool
+		wantContent   string
+		wantThinking  string
+		wantToolCalls int
+	}{
+		{
+			name:         "thinking enabled - simple thinking then content",
+			input:        "Let me think about this...</think>Here is my answer.",
+			thinkEnabled: true,
+			wantThinking: "Let me think about this...",
+			wantContent:  "Here is my answer.",
+		},
+		{
+			name:         "thinking enabled - only thinking",
+			input:        "I need to consider multiple factors...",
+			thinkEnabled: true,
+			wantThinking: "I need to consider multiple factors...",
+			wantContent:  "",
+		},
+		{
+			name:         "thinking disabled - direct content",
+			input:        "Here is my direct answer.",
+			thinkEnabled: false,
+			wantThinking: "",
+			wantContent:  "Here is my direct answer.",
+		},
+		{
+			name:          "thinking with tool call",
+			input:         "Let me search for that...</think>I'll use a tool.<tool_call>search<arg_key>query</arg_key><arg_value>test</arg_value></tool_call>",
+			thinkEnabled:  true,
+			wantThinking:  "Let me search for that...",
+			wantContent:   "I'll use a tool.",
+			wantToolCalls: 1,
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			p := &Parser{}
+
+			var thinkValue *api.ThinkValue
+			if tt.thinkEnabled {
+				thinkValue = &api.ThinkValue{Value: true}
+			} else {
+				thinkValue = &api.ThinkValue{Value: false}
+			}
+
+			// Define tools for tool call tests
+			props := api.NewToolPropertiesMap()
+			props.Set("query", api.ToolProperty{Type: api.PropertyType{"string"}})
+			tools := []api.Tool{
+				{
+					Function: api.ToolFunction{
+						Name: "search",
+						Parameters: api.ToolFunctionParameters{
+							Properties: props,
+						},
+					},
+				},
+			}
+
+			p.Init(tools, nil, thinkValue)
+
+			content, thinking, calls, err := p.Add(tt.input, true)
+			if err != nil {
+				t.Fatalf("unexpected error: %v", err)
+			}
+
+			if thinking != tt.wantThinking {
+				t.Errorf("thinking = %q, want %q", thinking, tt.wantThinking)
+			}
+			if content != tt.wantContent {
+				t.Errorf("content = %q, want %q", content, tt.wantContent)
+			}
+			if len(calls) != tt.wantToolCalls {
+				t.Errorf("len(calls) = %d, want %d", len(calls), tt.wantToolCalls)
+			}
+		})
+	}
+}
+
+func TestParserToolCall(t *testing.T) {
+	p := &Parser{}
+
+	props := api.NewToolPropertiesMap()
+	props.Set("location", api.ToolProperty{Type: api.PropertyType{"string"}})
+	props.Set("unit", api.ToolProperty{Type: api.PropertyType{"string"}})
+	tools := []api.Tool{
+		{
+			Function: api.ToolFunction{
+				Name: "get_weather",
+				Parameters: api.ToolFunctionParameters{
+					Properties: props,
+				},
+			},
+		},
+	}
+
+	// Initialize with thinking disabled
+	tv := &api.ThinkValue{Value: false}
+	p.Init(tools, nil, tv)
+
+	input := "<tool_call>get_weather<arg_key>location</arg_key><arg_value>San Francisco</arg_value><arg_key>unit</arg_key><arg_value>celsius</arg_value></tool_call>"
+
+	_, _, calls, err := p.Add(input, true)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+
+	if len(calls) != 1 {
+		t.Fatalf("expected 1 tool call, got %d", len(calls))
+	}
+
+	call := calls[0]
+	if call.Function.Name != "get_weather" {
+		t.Errorf("function name = %q, want %q", call.Function.Name, "get_weather")
+	}
+
+	location, ok := call.Function.Arguments.Get("location")
+	if !ok || location != "San Francisco" {
+		t.Errorf("location = %v, want %q", location, "San Francisco")
+	}
+
+	unit, ok := call.Function.Arguments.Get("unit")
+	if !ok || unit != "celsius" {
+		t.Errorf("unit = %v, want %q", unit, "celsius")
+	}
+}
+
+func TestOverlap(t *testing.T) {
+	tests := []struct {
+		s    string
+		tag  string
+		want int
+	}{
+		{"hello<", "</think>", 1},
+		{"hello</", "</think>", 2},
+		{"hello</t", "</think>", 3},
+		{"hello</th", "</think>", 4},
+		{"hello</thi", "</think>", 5},
+		{"hello</thin", "</think>", 6},
+		{"hello</think", "</think>", 7},
+		{"hello</think>", "</think>", 8}, // Complete tag at end returns full length
+		{"hello", "</think>", 0},
+		{"", "</think>", 0},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.s+"_"+tt.tag, func(t *testing.T) {
+			got := overlap(tt.s, tt.tag)
+			if got != tt.want {
+				t.Errorf("overlap(%q, %q) = %d, want %d", tt.s, tt.tag, got, tt.want)
+			}
+		})
+	}
+}
+
+func TestTrailingWhitespaceLen(t *testing.T) {
+	tests := []struct {
+		s    string
+		want int
+	}{
+		{"hello   ", 3},
+		{"hello\n\t ", 3},
+		{"hello", 0},
+		{"", 0},
+		{"   ", 3},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.s, func(t *testing.T) {
+			got := trailingWhitespaceLen(tt.s)
+			if got != tt.want {
+				t.Errorf("trailingWhitespaceLen(%q) = %d, want %d", tt.s, got, tt.want)
+			}
+		})
+	}
+}
--- a/x/models/glm4_moe_lite/render.go
+++ b/x/models/glm4_moe_lite/render.go
@@ -0,0 +1,173 @@
+package glm4_moe_lite
+
+import (
+	"encoding/json"
+	"fmt"
+	"strings"
+
+	"github.com/ollama/ollama/api"
+)
+
+// Renderer renders messages for GLM4-MoE-Lite models.
+//
+// GLM-4 Thinking Modes (ref: https://docs.z.ai/guides/capabilities/thinking-mode):
+//
+//  1. INTERLEAVED THINKING
+//     The model thinks between tool calls and after receiving tool results.
+//     This enables complex step-by-step reasoning: interpreting each tool output
+//     before deciding what to do next. Thinking blocks are preserved and returned
+//     with tool results to maintain reasoning continuity.
+//
+//  2. PRESERVED THINKING
+//     The model retains reasoning content from previous assistant turns in context.
+//     This preserves reasoning continuity across multi-turn conversations. The
+//     upstream API has a "clear_thinking" parameter to control this:
+//     - clear_thinking=true:  clears reasoning from previous turns (outputs </think>)
+//     - clear_thinking=false: preserves <think>...</think> blocks from previous turns
+//
+//  3. TURN-LEVEL THINKING
+//     Controls whether the model should reason on each turn. The upstream API
+//     uses "enable_thinking" parameter:
+//     - enable_thinking=true:  outputs <think> to start reasoning
+//     - enable_thinking=false: outputs </think> to skip reasoning
+//
+// OLLAMA DEFAULTS:
+//   - Thinking is ENABLED by default (thinkValue=nil or true outputs <think>)
+//   - Thinking is PRESERVED by default (reasoning content from previous turns is always
+//     included in <think>...</think> blocks, equivalent to clear_thinking=false)
+//   - Users can disable thinking per-turn via thinkValue=false
+type Renderer struct{}
+
+// Render renders messages into the GLM4 chat format.
+func (r *Renderer) Render(messages []api.Message, tools []api.Tool, thinkValue *api.ThinkValue) (string, error) {
+	var sb strings.Builder
+
+	sb.WriteString("[gMASK]<sop>")
+
+	if len(tools) > 0 {
+		sb.WriteString("<|system|>\n")
+		sb.WriteString("# Tools\n\n")
+		sb.WriteString("You may call one or more functions to assist with the user query.\n\n")
+		sb.WriteString("You are provided with function signatures within <tools></tools> XML tags:\n")
+		sb.WriteString("<tools>\n")
+		for _, tool := range tools {
+			d, _ := json.Marshal(tool)
+			sb.WriteString(formatToolJSON(d))
+			sb.WriteString("\n")
+		}
+		sb.WriteString("</tools>\n\n")
+		sb.WriteString("For each function call, output the function name and arguments within the following XML format:\n")
+		sb.WriteString("<tool_call>{function-name}<arg_key>{arg-key-1}</arg_key><arg_value>{arg-value-1}</arg_value><arg_key>{arg-key-2}</arg_key><arg_value>{arg-value-2}</arg_value>...</tool_call>")
+	}
+
+	think := true
+	if thinkValue != nil && !thinkValue.Bool() {
+		think = false
+	}
+
+	for i, message := range messages {
+		switch message.Role {
+		case "user":
+			sb.WriteString("<|user|>")
+			sb.WriteString(message.Content)
+		case "assistant":
+			sb.WriteString("<|assistant|>")
+			if message.Thinking != "" {
+				sb.WriteString("<think>" + message.Thinking + "</think>")
+			} else {
+				sb.WriteString("</think>")
+			}
+			if message.Content != "" {
+				sb.WriteString(message.Content)
+			}
+			if len(message.ToolCalls) > 0 {
+				for _, toolCall := range message.ToolCalls {
+					sb.WriteString("<tool_call>" + toolCall.Function.Name)
+					sb.WriteString(renderToolArguments(toolCall.Function.Arguments))
+					sb.WriteString("</tool_call>")
+				}
+			}
+		case "tool":
+			if i == 0 || messages[i-1].Role != "tool" {
+				sb.WriteString("<|observation|>")
+			}
+			sb.WriteString("<tool_response>")
+			sb.WriteString(message.Content)
+			sb.WriteString("</tool_response>")
+		case "system":
+			sb.WriteString("<|system|>")
+			sb.WriteString(message.Content)
+		}
+	}
+
+	sb.WriteString("<|assistant|>")
+	if think {
+		sb.WriteString("<think>")
+	} else {
+		sb.WriteString("</think>")
+	}
+
+	return sb.String(), nil
+}
+
+// renderToolArguments converts tool call arguments to GLM4 XML format.
+func renderToolArguments(args api.ToolCallFunctionArguments) string {
+	var sb strings.Builder
+	for key, value := range args.All() {
+		sb.WriteString("<arg_key>" + key + "</arg_key>")
+		var valueStr string
+		if str, ok := value.(string); ok {
+			valueStr = str
+		} else {
+			jsonBytes, err := json.Marshal(value)
+			if err != nil {
+				valueStr = fmt.Sprintf("%v", value)
+			} else {
+				valueStr = string(jsonBytes)
+			}
+		}
+
+		sb.WriteString("<arg_value>" + valueStr + "</arg_value>")
+	}
+
+	return sb.String()
+}
+
+// formatToolJSON formats JSON for GLM4 tool definitions by adding spaces after : and ,
+func formatToolJSON(raw []byte) string {
+	var sb strings.Builder
+	sb.Grow(len(raw) + len(raw)/10)
+
+	inString := false
+	escaped := false
+	for i := range raw {
+		ch := raw[i]
+		sb.WriteByte(ch)
+
+		if inString {
+			if escaped {
+				escaped = false
+				continue
+			}
+			if ch == '\\' {
+				escaped = true
+				continue
+			}
+			if ch == '"' {
+				inString = false
+			}
+			continue
+		}
+
+		if ch == '"' {
+			inString = true
+			continue
+		}
+
+		if ch == ':' || ch == ',' {
+			sb.WriteByte(' ')
+		}
+	}
+
+	return sb.String()
+}
--- a/x/models/glm4_moe_lite/render_test.go
+++ b/x/models/glm4_moe_lite/render_test.go
@@ -0,0 +1,203 @@
+package glm4_moe_lite
+
+import (
+	"strings"
+	"testing"
+
+	"github.com/ollama/ollama/api"
+)
+
+func TestRendererSimple(t *testing.T) {
+	r := &Renderer{}
+
+	messages := []api.Message{
+		{Role: "user", Content: "Hello"},
+	}
+
+	// Thinking enabled (default)
+	result, err := r.Render(messages, nil, nil)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+
+	expected := "[gMASK]<sop><|user|>Hello<|assistant|><think>"
+	if result != expected {
+		t.Errorf("result = %q, want %q", result, expected)
+	}
+}
+
+func TestRendererThinkingDisabled(t *testing.T) {
+	r := &Renderer{}
+
+	messages := []api.Message{
+		{Role: "user", Content: "Hello"},
+	}
+
+	tv := &api.ThinkValue{Value: false}
+
+	result, err := r.Render(messages, nil, tv)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+
+	expected := "[gMASK]<sop><|user|>Hello<|assistant|></think>"
+	if result != expected {
+		t.Errorf("result = %q, want %q", result, expected)
+	}
+}
+
+func TestRendererMultiTurn(t *testing.T) {
+	r := &Renderer{}
+
+	messages := []api.Message{
+		{Role: "user", Content: "What is 2+2?"},
+		{Role: "assistant", Content: "4", Thinking: "Let me calculate: 2+2=4"},
+		{Role: "user", Content: "And 3+3?"},
+	}
+
+	result, err := r.Render(messages, nil, nil)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+
+	// Check key parts
+	if !strings.Contains(result, "[gMASK]<sop>") {
+		t.Error("missing [gMASK]<sop> prefix")
+	}
+	if !strings.Contains(result, "<|user|>What is 2+2?") {
+		t.Error("missing first user message")
+	}
+	if !strings.Contains(result, "<|assistant|><think>Let me calculate: 2+2=4</think>4") {
+		t.Error("missing assistant message with thinking")
+	}
+	if !strings.Contains(result, "<|user|>And 3+3?") {
+		t.Error("missing second user message")
+	}
+	if !strings.HasSuffix(result, "<|assistant|><think>") {
+		t.Errorf("should end with <|assistant|><think>, got suffix: %q", result[len(result)-30:])
+	}
+}
+
+func TestRendererWithSystem(t *testing.T) {
+	r := &Renderer{}
+
+	messages := []api.Message{
+		{Role: "system", Content: "You are a helpful assistant."},
+		{Role: "user", Content: "Hello"},
+	}
+
+	result, err := r.Render(messages, nil, nil)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+
+	if !strings.Contains(result, "<|system|>You are a helpful assistant.") {
+		t.Error("missing system message")
+	}
+}
+
+func TestRendererWithTools(t *testing.T) {
+	r := &Renderer{}
+
+	messages := []api.Message{
+		{Role: "user", Content: "What's the weather?"},
+	}
+
+	props := api.NewToolPropertiesMap()
+	props.Set("location", api.ToolProperty{Type: api.PropertyType{"string"}, Description: "The city"})
+	tools := []api.Tool{
+		{
+			Function: api.ToolFunction{
+				Name:        "get_weather",
+				Description: "Get the weather for a location",
+				Parameters: api.ToolFunctionParameters{
+					Type:       "object",
+					Properties: props,
+					Required:   []string{"location"},
+				},
+			},
+		},
+	}
+
+	result, err := r.Render(messages, tools, nil)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+
+	// Check for tool system prompt
+	if !strings.Contains(result, "<|system|>") {
+		t.Error("missing system tag for tools")
+	}
+	if !strings.Contains(result, "# Tools") {
+		t.Error("missing tools header")
+	}
+	if !strings.Contains(result, "<tools>") {
+		t.Error("missing tools tag")
+	}
+	if !strings.Contains(result, "get_weather") {
+		t.Error("missing tool name")
+	}
+	if !strings.Contains(result, "</tools>") {
+		t.Error("missing closing tools tag")
+	}
+}
+
+func TestRendererWithToolCalls(t *testing.T) {
+	r := &Renderer{}
+
+	args := api.NewToolCallFunctionArguments()
+	args.Set("location", "San Francisco")
+
+	messages := []api.Message{
+		{Role: "user", Content: "What's the weather in SF?"},
+		{
+			Role: "assistant",
+			ToolCalls: []api.ToolCall{
+				{
+					Function: api.ToolCallFunction{
+						Name:      "get_weather",
+						Arguments: args,
+					},
+				},
+			},
+		},
+		{Role: "tool", Content: "Sunny, 72F"},
+	}
+
+	result, err := r.Render(messages, nil, nil)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+
+	if !strings.Contains(result, "<tool_call>get_weather") {
+		t.Error("missing tool call")
+	}
+	if !strings.Contains(result, "<arg_key>location</arg_key>") {
+		t.Error("missing arg_key")
+	}
+	if !strings.Contains(result, "<arg_value>San Francisco</arg_value>") {
+		t.Error("missing arg_value")
+	}
+	if !strings.Contains(result, "</tool_call>") {
+		t.Error("missing tool call closing tag")
+	}
+	if !strings.Contains(result, "<|observation|>") {
+		t.Error("missing observation tag")
+	}
+	if !strings.Contains(result, "<tool_response>Sunny, 72F</tool_response>") {
+		t.Error("missing tool response")
+	}
+}
+
+func TestFormatToolJSON(t *testing.T) {
+	input := []byte(`{"name":"test","value":123}`)
+	result := formatToolJSON(input)
+
+	// Should add spaces after : and ,
+	if !strings.Contains(result, ": ") {
+		t.Error("should add space after colon")
+	}
+	if !strings.Contains(result, ", ") {
+		t.Error("should add space after comma")
+	}
+}