You can deploy this model and start earning money today!

majentik/
Qwen3.6-35B-A3B-FP8

Name: Qwen3.6-35B-A3B-FP8
Availability: InStock

Qwen/Qwen3.6-35B-A3B-FP8

finetune

apache-2.0

endpoints_compatible

Views

No views yet

[!NOTE] Status (2026-07-07): no weights published yet. This repository currently contains only the model card — it marks a planned variant that has not been released. Follow the repo to be notified when files land.

Qwen3.6-35B-A3B-FP8

Summary

Reference wrapper around Qwen/Qwen3.6-35B-A3B-FP8 — the official FP8 release. This repository carries no weights; it exists only to anchor the FP8 variant inside the majentik/* family navigation.

Why this variant

Pick this for Hopper / Ada / Blackwell GPUs where FP8 is natively supported and you want the closest-to-bf16 fidelity with ~50% memory savings. For additional compression pick one of the 4-bit variants below.

Hardware compatibility

Device	VRAM	Recommendation
H100 / H200	80–141 GB	native
RTX 4090	24 GB	does not fit full precision — use 4-bit
RTX 5090	32 GB	native

Reproduce

bash

# No re-quantization needed — use the upstream weights directly.
huggingface-cli download Qwen/Qwen3.6-35B-A3B-FP8

Evaluation

This FP8 pack is a weights + config mirror and is not loadable by mlx_lm, so it is not covered by the local MLX eval harness. For measured scores on this family, see the benchmark tables on the sibling MLX variants (e.g. qwen3.6-35b-a3b-mlx-mxfp4 and qwen3.6-35b-a3b-mlx-nvfp4), which were evaluated with mlx_lm.evaluate (arc_easy, hellaswag; limit 200).

Family

bf16 — Qwen/Qwen3.6-35B-A3B
FP8 (this) — Qwen/Qwen3.6-35B-A3B-FP8

Provenance

Card-only. No weights stored.

License

Released under apache-2.0. Upstream license of the base model applies.

majentik/Qwen3.6-35B-A3B-FP8