Qwen3.8-27B-NVFP4A16

Status: placeholder. There are no weights in this repository yet.

Qwen/Qwen3.8-27B has not been released at the time of writing. This repo is a reserved skeleton for a planned quantisation; it exists so the naming and the intended recipe are public up front. Do not attempt to download or serve this repo — there is nothing here but this card. When the base model ships and the quantisation has been built and measured, this card will be replaced with real weights, a recipe, and KLD/perplexity numbers.

Planned build

• Base model: Qwen/Qwen3.8-27B (unreleased)
• Scheme: NVFP4A16 (W4A16), attention kept in bf16
• Tool: llm-compressor, compressed-tensors format
• Recipe: recipes/nvfp4a16_attnbf16.yaml in the pipeline repo

NVFP4 4-bit weights with bf16 activations, over the LLM trunk's MLPs. The vision tower, the linear_attn SSM block, the entire self_attn block and the MTP head stay in bf16. W4A16 rather than W4A4 — on every Qwen3.5/3.6-family base we have measured, activation quantisation roughly doubles KLD for zero footprint benefit (the two are the same size on disk).

What is not yet known

The recipe above is the intended configuration, carried over from our Qwen3.6-27B-family builds. It is contingent on Qwen/Qwen3.8-27B having a comparable architecture. If the release turns out to be mixture-of-experts, to use a different module layout, or to ship without the multi-token-prediction head, the recipe will change and this card will say so.

No quality numbers are claimed here. Every published build in this line ships with per-token KL divergence against the bf16 base and wikitext-2-raw perplexity, measured on our own harness; those will be added when they exist.

Provenance

Built by huginnfork. This is a plain quantisation — no abliteration, no fine-tuning, no merging. Weights derive solely from the upstream base model, whose license and terms apply.