mlx-community/Xing4.0-29B-A4B-OptiQ-4bit

Built with mlx-optiq, the MLX-native toolkit to quantize, fine-tune, and serve LLMs locally on Apple Silicon, no PyTorch and no cloud. Family guide · All OptiQ quants · Docs

A mixed-precision MLX quant of XingChen-AGI/Xing4.0-29B-A4B. Sensitive layers are kept at 8-bit and robust ones at 4-bit, rather than crushing everything to a uniform width.

Requires mlx-optiq

Stock mlx-lm has no xing4_0 class — support is an unmerged pull request, so a plain install answers ValueError: Model type xing4_0 not supported. OptiQ ships a vendored, mlx-native port that registers with mlx-lm on import optiq.

pip install "mlx-optiq>=0.5.13"

Use it

optiq serve --model mlx-community/Xing4.0-29B-A4B-OptiQ-4bit

Or with mlx-lm directly:

import optiq  # registers the xing4_0 arch with mlx-lm
from mlx_lm import load, generate

model, tok = load("mlx-community/Xing4.0-29B-A4B-OptiQ-4bit")
prompt = tok.apply_chat_template(
    [{"role": "user", "content": "Explain write-ahead logging in two sentences."}],
    add_generation_prompt=True, tokenize=False)
print(generate(model, tok, prompt=prompt, max_tokens=256))

Quantization details

Property	Value
Predominant precision	4-bit
Layers at 8-bit	303
Bits per weight	5.00
Size on disk	19.6 GB (from ~58 GB bf16)
Group size	64

The per-layer allocation was measured on this model: a KL-divergence sensitivity pass over all 395 quantizable tensors — the 64 routed experts included — against a six-domain calibration mix, then a knapsack over the bit budget. The per-layer bit map is in optiq/metadata.json and in the quantization block of config.json.

Capability Score

Measured on this quant, not inherited from the parent. Xing4.0 is a new family with no scored sibling, so the full six-metric suite was run.

Metric	Score
MMLU (5-shot, 1000 samples)	71.7%
GSM8K (1000 samples)	90.1%
IFEval (strict, prompt-level)	61.2%
BFCL-V3 simple	74.5%
HumanEval pass@1	84.8%
HashHop (~8.9k context)	18.0%
Capability Score	66.71

Read the HashHop row in context. 18% on long-context hash retrieval sits far below the rest, and it is the architecture rather than the quantization: multi-head latent attention compresses the KV cache into a low-rank latent (kv_lora_rank: 512), which is what needle-in-a-haystack recall over thousands of random keys pays for. The model answers immediately with a well-formed but wrong hash — it is not running out of tokens and it is not being mis-scored. Reasoning, instruction-following, tool use and coding are unaffected.

The architecture

DeepSeek-V3's attention — multi-head latent attention, 64 routed experts and one shared expert, about 4B parameters active per token — plus one thing of its own: the hyper-connection. Instead of a single residual stream running through the network, it carries four in parallel, and every block reads from and writes back to a mix of them decided per layer by a Sinkhorn-normalised routing matrix.

Notes

• A bf16 reference comparison was not possible on the machine that built this (the parent is 58 GB against 36 GB of RAM), so the port was verified by a line-by-line reading against the reference implementation plus generation against the quant, not by a bit-exact check.