Qwen 3.6 35B-A3B — JANGTQ2 (MLX)
TurboQuant codebook quantization of Alibaba's hybrid linear/full-attention agentic MoE — routed experts at 2-bit via Lloyd-Max codebooks + Hadamard rotation, attention / embed / shared-expert / lm_head at 8-bit affine, vision tower preserved.
Model Details
| Property | Value |
|---|
| Base model | Qwen/Qwen3.6-35B-A3B |
| Parameters (source) | 35 B total, ~3 B active per token |
| Architecture | qwen3_5_moe — 40 decoder layers: 30 Gated DeltaNet (linear attn) + 10 full attention, 256 routed experts + 1 always-on shared expert |
| Quantization format | weight_format: mxtq — routed experts via TurboQuant codebook (2-bit), everything else affine 8-bit or fp16 passthrough |
| Routed-expert storage | .tq_packed (uint32) + .tq_norms (fp16) + .tq_bits (uint8); codebook + Hadamard signs re-derived deterministically at load |
| Package size on disk | 11.63 GB across 12 shards |
| Shipped tensors | 1,930 total (1,597 language-model + 333 vision tower + 120 routed-expert TQ triples) |
| Vocab | 248,320 |
| Context (position embeddings) | 262,144 native; the upstream model card reports up to ~1 M with YaRN scaling |
| Vision tower | 27-layer ViT (hidden 1152, patch 16), preserved in fp16 |
| Chat format | Qwen im_start/im_end, unified thinking toggle |
Quantization details, per tensor category
| Category | Bits | Group / codebook | Notes |
|---|
Routed-expert MLP (mlp.experts.gate_up_proj, down_proj) | 2 (JANGTQ) | 2^2 Lloyd-Max centroids + Hadamard rotation | .tq_packed + .tq_norms + .tq_bits triples |
Embedding (embed_tokens), lm_head | 8 (affine) | group 64 | MLX-native QuantizedLinear |
Full-attention projections (q_proj, k_proj, v_proj, o_proj) | 8 (affine) | group 64 | Gate-doubled q_proj for attn_output_gate |
Linear-attention projections (in_proj_qkv, in_proj_z, in_proj_b, in_proj_a, out_proj) | 8 (affine) | group 64 | Gated DeltaNet |
Shared-expert MLP (gate_proj, up_proj, down_proj) | 8 (affine) | group 64 | Always active per token |
Router (mlp.gate) | fp16 passthrough | — | Precision-critical |
Shared-expert gate (shared_expert_gate) | fp16 passthrough | — | sigmoid scalar gate |
Norms (*_layernorm, *_norm), A_log, dt_bias, conv1d | fp16 passthrough | — | Un-quantized |
| Vision tower (333 tensors) | fp16 passthrough | — | patch_embed.proj axes pre-transposed to MLX layout |
JANGTQ ("TurboQuant") stores routed-expert weights as indices into a small Lloyd-Max codebook with a per-row norm, after a randomized Hadamard rotation that concentrates the distribution so quantization error is uniform. At inference, the input is rotated once per layer (cheap fused Metal kernel) and dot products happen against the codebook centroids directly, so we never dequantize back to affine. Compared to affine 2-bit at the same bit budget, this gives better quality AND faster decode on the routed-expert MLP path.
Usage
JANGTQ requires our custom loader — stock
mlx_lm.load() can't parse
.tq_packed tensors. You need
jang-tools (free, public):
https://github.com/jjang-ai/jangq.
1pip install mlx mlx-lm mlx-vlm
2git clone https://github.com/jjang-ai/jangq && pip install -e ./jangq/jang-tools
Text
1from jang_tools.load_jangtq import load_jangtq_model
2from mlx_lm import generate
3
4model, tokenizer = load_jangtq_model("OsaurusAI/Qwen3.6-35B-A3B-JANGTQ2")
5print(generate(model, tokenizer,
6 prompt="The capital of France is",
7 max_tokens=64))
Image (VLM)
1from jang_tools.load_jangtq_vlm import load_jangtq_vlm_model
2from mlx_vlm import generate
3from mlx_vlm.prompt_utils import apply_chat_template
4from mlx_vlm.utils import load_config
5
6path = "OsaurusAI/Qwen3.6-35B-A3B-JANGTQ2"
7model, processor = load_jangtq_vlm_model(path)
8config = load_config(path)
9
10prompt = apply_chat_template(processor, config, "Describe this image.", num_images=1)
11print(generate(model, processor, prompt, image="path/to/image.jpg", max_tokens=200))
Reasoning toggle
1msgs = [{"role": "user", "content": "What is 17 × 23?"}]
2# Reasoning OFF — pre-closed <think></think> block
3prompt = tokenizer.apply_chat_template(msgs, add_generation_prompt=True,
4 enable_thinking=False)
5# Reasoning ON — model fills the <think> block
6prompt = tokenizer.apply_chat_template(msgs, add_generation_prompt=True,
7 enable_thinking=True)
Pass enable_thinking as a direct kwarg (the chat_template_kwargs={...} form only propagates on some tokenizer versions).
Video
The base model supports video via transformers and the bundle preserves video_preprocessor_config.json. mlx-vlm 0.4.4's prepare_inputs has no video path yet for qwen3_5_moe — the Python load_jangtq_vlm path wraps video via a custom processor for our test harness. For mainline mlx-vlm users, stick to image input; use upstream transformers for video.
Hardware notes
~12 GB on disk; expect ~12–14 GB resident after load, plus KV cache.
| Mac unified RAM | Works? | Notes |
|---|
| 24 GB | ✅ comfortable | Full 32 k context OK |
| 32 GB | ✅ | 32-100 k context depending on profile |
| 24 GB | ✅ | text-only, short context |
Benchmarks
Base-model reference (Qwen/Qwen3.6-35B-A3B, upstream, not this quant):
| MMLU-Pro | AIME 2026 | LiveCodeBench v6 | GPQA | SWE-bench Verified |
|---|
| 85.2 | 92.7 | 80.4 | 86.0 | 73.4 |
Independent JANGTQ-quant evaluation is tracked in the jang-tools repo and will land in future README revisions.
Citation
1@misc{qwen2026qwen36,
2 title = {Qwen3.6-Plus: Towards Real World Agents},
3 author = {Qwen Team},
4 year = {2026},
5 url = {https://qwen.ai/blog?id=qwen3.6}
6}
License
Apache 2.0 — inherits from the base model.
Packaged on Apple Silicon with
jang-tools (mlx-lm 0.31.2).
© 2026 Osaurus AI —
osaurus.ai
