mlx-community

/

Qwen3-Coder-Next-4bit

like 8

Follow

MLX Community 8.91k

Summary

I'm encountering a model loading error with Qwen3-Coder-Next-4bit that appears similar to the pixtral-12b-8bit quantization issue reported previously.

Error

ValueError: Received 1084 parameters not in model:
lm_head.biases,
lm_head.scales,
model.embed_tokens.biases,
model.embed_tokens.scales,
model.layers.0.linear_attn.in_proj_ba.biases,
model.layers.0.linear_attn.in_proj_ba.scales,
[... and 1078 more similar .biases/.scales parameters ...]

The server crashes at layer 9 during model loading. All 9 safetensor files are present and complete (~43GB total).

Reproduction

python3 -m mlx_lm.server \
  --model /path/to/qwen3-coder-next-80b-a3b-moe-coder-2507-4bit-mlx-sft-256k-exp512 \
  --port 8201 \
  --host 127.0.0.1 \
  --max-tokens 2048

Server loads 8 layers successfully, then crashes with the ValueError above.

Environment

• Model: mlx-community/Qwen3-Coder-Next-4bit (qwen3-coder-next-80b variant)
• MLX Version: 0.30.5 (from model README)
• Model Size: ~43GB (512 experts × 48 layers MoE)
• Platform: macOS, Apple Silicon
• Model Date: Downloaded Feb 3-7, 2026 (very recent)

Analysis

This looks like the kind of problem from the pixtral-12b-8bit issue where quantization parameters (.biases and .scales) were included in the model files but not expected by the MLX architecture. The error shows 1084 parameters across all layers that have these quantization artifacts.

I'm not certain if this is the exact same root cause, but the symptoms appear very similar:

• Quantization-related parameters (.biases, .scales)
• Parameters present in safetensors but not in model architecture
• Recent model conversion (mlx-lm 0.30.5)

Question

Is this a known issue with the Qwen3-Coder-Next conversion? I couldn't find any other reports of this specific problem. If it's the same quantization bug that affected pixtral, is there a timeline for a fix?

I could be mistaken about the root cause - any guidance would be appreciated.

Files Verified

All expected files are present:

• ✅ config.json (512 experts, 48 layers)
• ✅ model-00001-of-00009.safetensors through model-00009-of-00009.safetensors
• ✅ tokenizer.json, tokenizer_config.json
• ✅ README.md (shows mlx-lm 0.30.5 conversion)

Thank you for maintaining these MLX conversions!