Hy3-preview-JANGTQ

<p align="center"><img src="jangq-logo.png" width="160"/></p> <p align="center"> <a href="https://vmlx.net"><img src="vmlx-app.png" alt="vMLX — run JANG models on Apple Silicon" width="820"></a> </p> <h3 align="center">⚡ All JANG models are meant to be run in <a href="https://vmlx.net">vMLX</a></h3>

Hy3-preview-JANGTQ

Tencent Hy3-preview — 79 GB on disk (down from the ~557 GB BF16 source) — 2-bit JANGTQ quantization on routed experts + 8-bit affine elsewhere.

• Source: tencent/Hy3-preview (Hy3 architecture, 295B total / 21B active, BF16 native, 256K context, 80 transformer layers + 1 MTP, 192 routed experts top-8 + 1 shared)
• Quantization: JANGTQ — 2-bit MXTQ codebook (Hadamard-rotated, Lloyd-Max optimized) on routed-expert weights + 8-bit affine on attention / shared expert / dense layer-0 / embed / lm_head / MTP matmuls + fp16 passthrough on RMSNorms / router gate / expert_bias
• MTP: layer 80 weights preserved (mtp_mode=preserved_disabled); decode is one-token-per-forward until accept/reject speculative loop ships
• Bundle size: 79 GB on-disk across 85 shards
• Runs on: M4 Max 128 GB / M5 Max 128 GB / Mac Studio 192 GB+

What's in the bundle

jangtq_runtime.safetensors sidecar (~22 KB) for Swift runtimes — covers (in_features={1536, 4096}, seed=42, bits=2) codebooks + sign-flip vectors.

Loading (Python)

pip install jang-tools mlx-lm

from jang_tools.load_jangtq import load_jangtq_model

model, tokenizer = load_jangtq_model("JANGQ-AI/Hy3-preview-JANGTQ")

chat = tokenizer.apply_chat_template(
    [{"role": "user", "content": "What is 2 + 2? Answer briefly."}],
    tokenize=False,
    add_generation_prompt=True,
    reasoning_effort="no_think",
)

load_jangtq_model auto-registers model_type=hy_v3 via jang_tools.hy3 before building the MLX skeleton. The loader applies the standard SwitchGLU fused gate+up + P15 router compile + P18 QKV fusion patches automatically. Two Hy3-specific runtime fixes are baked in:

1. fp32 lm_head. enable_lm_head_fp32=True in the bundle config — Model.__call__ dequantizes the quantized lm_head and accumulates the 4096-dim contraction in fp32 (mirrors DSV4's pattern). bf16 accumulation drifts logits by ~0.5/elem and flips top-k token picks toward high-baseline-energy junk tokens.
2. qk_norm under JANGTQ P18 QKV fusion. JANGTQ's QKV-fusion patch replaces the attention __call__; Hy3Attention declares use_qk_norm=True and uses Hy3HeadRMSNorm to auto-reshape flat [B, L, n_heads * head_dim] input to per-head shape so RMSNorm normalizes over head_dim, not over the entire flat dimension.

Decode ~15 tok/s greedy on M5 Max 128 GB at reasoning_effort=no_think.

Reasoning + tools

• Reasoning parser: qwen3 (extracts <think>...</think> blocks)
• Tool parser: hunyuan (Tencent XML-like: <tool_calls><tool_call>name<tool_sep><arg_key>k</arg_key><arg_value>v</arg_value></tool_call></tool_calls>)
• Reasoning effort: no_think (default) | low | high — pass via apply_chat_template(..., reasoning_effort="…")
• Default rendering: template emits a closed <think></think> for no_think mode; the runtime should NOT auto-open a reasoning prefix unless low or high is explicitly requested
• Cache: kv (standard GQA cache; no MLA, no SSM, no sliding-window)

Top-K runtime override

JANGTQ_TOPK_OVERRIDE=4 python serve.py lowers per-token expert count from the trained 8 to 4 for ~10% decode speedup. Coherence holds on short prompts in our smoke tests; long-form quality is not benchmarked. The patcher refuses to set K above the trained value and logs the attribute count it modified.

Credits

• Quantization + MLX runtime: Jinho Jang ([email protected])
• Source model: Tencent Hy3-preview team
• License: Tencent Hy Community License — non-commercial, EU/UK/SK excluded; consult the LICENSE for full terms

Validated runtime contract

• 80 layers materialize; 79 routed-expert SwitchGLU instances hydrate via TurboQuantLinear (2-bit MXTQ).
• Capabilities verify: family=hy_v3, reasoning_parser=qwen3, tool_parser=hunyuan, think_in_template=False, supports_thinking=True, cache_type=kv, modality=text.
• Coherence smoke (M5 Max 128 GB):
  • "What is 2 + 2?" → 4<｜hy_eos｜> (15.2 tok/s)
  • "The capital of France is" → top-1 Paris (logit 19.13)
  • "def fibonacci(n):" → top-1 \n, top-3 includes return
• Hard-prompt benchmark coverage (HumanEval, MMLU, long-context) is pending. This bundle is shipped on smoke evidence; treat results beyond short prompts as preview-quality until benchmarks land.

Runtime support matrix