Modelpublic

LibraxisAI/Huihui-Qwen3.6-35B-A3B-Claude-4.7-Opus-abliterated-vmlx-mxfp8

sourceHugging Faceapache-2.0updated 5mo agoView on Hugging Face

1likes165downloads

Model Card

Huihui-Qwen3.6-35B-A3B-Claude-4.7-Opus-abliterated-vmlx-mxfp8

Huihui-Qwen3.6-35B-A3B-Claude-4.7-Opus-abliterated-vmlx-mxfp8 is an MLX / VMLX vision-language checkpoint derived from huihui-ai/Huihui-Qwen3.6-35B-A3B-Claude-4.7-Opus-abliterated, packaged for local multimodal experimentation on Apple Silicon.

Tested inference path

Inference for this checkpoint has been tested with [`LibraxisAI/mlx-batch-server`](https://github.com/LibraxisAI/mlx-batch-server). This is the recommended tested path for operator-controlled local multimodal mlx-lm inference on Apple Silicon.

Aspect	Status
Tested runtime	`LibraxisAI/mlx-batch-server`
Target hardware	Apple Silicon
Inference mode	Local / self-hosted
Hugging Face Hosted Inference	Disabled for this repository (`inference: false`)

This does not claim compatibility with every possible serving stack. It documents the path that has been exercised for this published checkpoint.

Intended use

—Local image-and-text reasoning on Apple Silicon
—Multimodal prompting experiments
—Screenshot, document, chart, and visual question-answering workflows
—Operator-controlled local inference where hosted inference is not desired

Out of scope

—Safety-critical decisions without domain expert review
—Claims of benchmark superiority not backed by published evaluation data
—Non-MLX / non-VMLX runtime guarantees
—High-stakes visual interpretation without human validation

Training and conversion metadata

Parameter	Value
Repository	`LibraxisAI/Huihui-Qwen3.6-35B-A3B-Claude-4.7-Opus-abliterated-vmlx-mxfp8`
Base model	`huihui-ai/Huihui-Qwen3.6-35B-A3B-Claude-4.7-Opus-abliterated`
Task	`image-text-to-text`
Library	`mlx`
Format	MLX / VMLX checkpoint
Quantization	MXFP8
Target platform	Apple Silicon

This card reports metadata present in the Hugging Face repository, existing frontmatter, or public config files. Missing benchmark, dataset, or training-run details are left explicit rather than reconstructed.

Usage

Use the library instructions above, or run this checkpoint through the tested local serving path: `LibraxisAI/mlx-batch-server`

Validation

End-to-end pipeline test 2026-04-22 on M3 Ultra (load → text → vision → unload), served via mlx-batch-server:

Probe	TTFT	Output chars	Notes
Cold load	—	—	39 s from cold to ready
Text — simple greeting (PL)	0.75 s	438	Clean output, abliterated behaviour
Text — canonical (PL, literary)	0.37 s	690	Concise reasoning trace
Vision — JPEG (Monument Valley)	13.14 s	1149	Detailed scene description

3/3 probes passed. has_reasoning=True on all probes — this model emits reasoning traces via <think> markers.

Limitations

—Validate outputs on your own domain data before relying on this checkpoint.
—Memory use and speed depend heavily on Apple Silicon generation, unified-memory size, prompt length, and runtime configuration.
—Validation data above reflects M3 Ultra; expect different timings on other hardware.

License

apache-2.0. Check the upstream/base model license as well when a base model is declared.