qwopus3.6-35b-a3b-coder-mxfp8-mlx

MLX MXFP8 MLX conversion of Jackrong/Qwopus3.6-35B-A3B-Coder, prepared by Shiftedx for Apple Silicon / MLX / LM Studio.

What Changed

Converted from the upstream safetensors checkpoint with the local streaming MLX pipeline.
Quantized primary linear weights with mxfp8 at group size 32.
Kept MoE router/gate modules in affine 8-bit group size 64 for compatibility.
Removed source MTP tensors and set MTP/next-token prediction layer counts to 0 for LM Studio compatibility.
Set tool_parser_type to qwen3_coder.
Patched the chat template so enable_thinking defaults to false when a runtime honors the template variable.
Removed vision tensors/config for a smaller language-only build.

Local Validation

Hardware reference: Apple M4 Max Apple Silicon with 64 GB unified memory.

Validated locally on June 29, 2026 with LM Studio server on port 8080, 32k context, parallel 1, GPU max.

Check	Result
LM Studio load	Passed; 33.31 GiB in LM Studio at 32k context.
Basic text completion	Passed; returned `2+2=4` and stopped.
Vision image smoke	Not applicable; this is the language-only build.

Note: LM Studio may still report hidden reasoning_tokens for this checkpoint even though the upstream model is intended for thinking-off use. Use adequate max_tokens for smoke tests.

LM Studio

After downloading in LM Studio, load the model key:

lms load qwopus3.6-35b-a3b-coder-mxfp8-mlx --context-length 32768 --parallel 1 --gpu max

Recommended profile defaults, matching the local Shiftedx 35B AgentWorld/Ornith profiles:

Preset/template: Qwen3 thinking-compatible Jinja template with <|im_end|> stop.
Context length: 200000 when memory allows; 32768 was used for local smoke validation.
Sampling: temperature 0.6, top-k 20, top-p 0.95, min-p enabled at 0.
Repeat penalty: unchecked/off, value 1.1 if enabled manually.
Load: parallel 1, GPU max.

Provenance

Source model: https://huggingface.co/Jackrong/Qwopus3.6-35B-A3B-Coder
Source license: Apache-2.0
Quantization date: June 29, 2026