Qwen Concept 27B

Qwen3.8-27B with nine down-projections rewritten by weight-space programming. No gradient steps, no training data: the edit is derived from the model's own activations and compiled into nine weight matrices.

The model thinks less and scores higher. Against the unmodified base, on a pre-registered criterion: GPQA Diamond +4.3 [+1.8, +7.1], LiveCodeBench v6 +13.0, with thinking tokens down 27-54%.

The contribution is the control, not the edit. A norm-matched random edit — same layers, same magnitude, different direction — does nothing, and its confidence intervals straddle zero on every benchmark. That is what separates "this direction means something" from "perturbation helps".

Paper, data and the code that generates every number: https://concept.country/research/weight-space-programming

Formats

folder	format	size	notes
`bf16/`	bfloat16	~52 GB	full precision; every reported number comes from these weights
`fp8/`	FP8_BLOCK	~30 GB	matches Qwen's official FP8 layout
`nvfp4a16/`	NVFP4 W4A16	~20 GB	NVFP4 weights, BF16 activations; Blackwell-class hardware
`gguf/`	Q8_0	~27 GB	llama.cpp; the exact file the benchmarks ran on
`gguf/`	Q4_K_M	~16 GB	llama.cpp; quantized one hop from bf16, not from the Q8_0

The quantizations are deployment conveniences, not separately evaluated artifacts. Every number above is from bf16 weights served at Q8_0. Nothing here claims a 4-bit conversion preserves the effect — that needs its own evaluation and has not had one.

Each folder carries a SHA256SUMS file.

Also available

Mirrored on the Concept Country registry, which speaks the HF Hub API:

HF_ENDPOINT=https://models.concept.country hf download ConceptCountry/qwen-concept-27b

Status

Experimental, under evaluation.

License

Apache-2.0, inherited from Qwen3.8-27B (Alibaba Cloud).