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

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

Build Notes

Quantized primary language weights with mxfp8 at group size 32.
Kept MoE router and gate modules in affine 8-bit group size 64 for compatibility.
Added Qwen3.5/Qwen3.6-MoE-compatible vision components and validated image grounding locally.
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 the runtime honors that template variable.
Applied a research refusal-direction weight edit using residual-direction orthogonalization.

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

Validated locally on June 30, 2026 and July 1, 2026 with LM Studio and direct MLX/VLM loading.

Check	Result
LM Studio load	Passed at 32k context, parallel 1, GPU max.
Basic text completion	Passed; answered `2+2` with `4` and stopped.
Code completion	Passed; produced a simple valid `add(a, b)` function.
Direct MLX/VLM image color smoke	Passed; answered `Red`.
Direct MLX/VLM OCR smoke	Passed; answered `FABLE 42`.
LM Studio OpenAI-compatible image smoke	Passed; answered `Red` and `FABLE 42`.
LM Studio native image smoke	Passed; answered `Red` and `FABLE 42`.
Thinking-off behavior	Smoke checks returned `0` reasoning tokens.
LM Studio logs	No warnings, errors, tracebacks, KV-cache issues, or tokenizer-regex warnings in the validation window.

This is a smoke-validated release, not a full benchmark suite. Broader downstream evaluation is still recommended for production use.

After downloading in LM Studio, load the model by repo name:

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

Recommended profile defaults:

Preset/template: Qwen3 thinking-compatible Jinja template with <|im_end|> stop.
Thinking: off by default through the included chat template.
Context length: 200000 when memory allows; 32768 was used for local validation.
Sampling: temperature 0.6, top-k 20, top-p 0.95, min-p enabled at 0.
Repeat penalty: off by default.
Load: parallel 1, GPU max.