Request access to the BottleCap AI model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Request access below. Add your company if you're evaluating this for work — we have enterprise versions that go further, and we'll make sure you hear about them first. We'll also send you new ThinkingCap releases and early access before they're public.

Tell us how you plan to use the model and we can help you get the most out of it — there's a short form for that too.

ThinkingCap-Qwen3.8-27B-NVFP4

NVFP4 weight-only (NVFP4A16, compressed-tensors nvfp4-pack-quantized, 16-element FP4 groups with FP8 scales), built by llm-compressor. The two GDN projections vLLM fuses into one GEMM (in_proj_qkv, in_proj_z) share one global scale: vLLM keeps a single scale for the fused GEMM, so separate scales would dequantize one of the two too small. Runs on Hopper (Marlin) and Blackwell.

Built from bottlecapai/ThinkingCap-Qwen3.8-27B (bf16). Vision tower, MTP head, lm_head and the GDN in_proj_a / in_proj_b projections stay bf16 (recipe.yaml). Serve with vLLM 0.29 (--trust-remote-code not needed).

Expected performance

Accuracy and completion length against the bf16 weights on the full quantization plan. Both rows are served by vLLM and answer the same questions — MMLU-Pro 1,500, RealWorldQA 765 × 2 seeds (images), GPQA-Diamond 198 × 4 seeds, IFBench 300 × 2 seeds, AA-LCR 100 (long-context prompts) — thinking at the chat template's default reasoning effort (xhigh), with sampled decoding (temperature 1.0, top_p 0.95, top_k 20, min_p 0.0) and a 65,536-token generation cap. acc is the mean over seeds, and the ± is the 95% interval over questions (each question averaged over its seeds first). AA-LCR answers are graded by Gemma-4-26B-A4B-it. median tokens / mean tokens = completion length (reasoning plus answer) over the questions; with several seeds, the median and mean of the per-seed values.

NVFP4 stays within that interval of bf16 on every benchmark; the largest gap is AA-LCR, 3.0 points lower on 100 questions. Evaluated, not proven lossless: the intervals leave room for differences of a few points.

MMLU-Pro (reasoning) — 1500 questions × 1 seed

config	acc (mean ± 95% CI)	median tokens	mean tokens
ThinkingCap-Qwen3.8-27B bf16	0.841 ± 0.018	166	1436
NVFP4	0.851 ± 0.018	170	1330

RealWorldQA (vision) — 765 questions × 2 seeds

config	acc (mean ± 95% CI)	median tokens	mean tokens
ThinkingCap-Qwen3.8-27B bf16	0.831 ± 0.024	112	488
NVFP4	0.825 ± 0.025	116	461

GPQA-Diamond (graduate-level science) — 198 questions × 4 seeds

config	acc (mean ± 95% CI)	median tokens	mean tokens
ThinkingCap-Qwen3.8-27B bf16	0.880 ± 0.037	1045	7115
NVFP4	0.864 ± 0.038	1136	6618

IFBench (instruction following) — 300 questions × 2 seeds

config	acc (mean ± 95% CI)	median tokens	mean tokens
ThinkingCap-Qwen3.8-27B bf16	0.797 ± 0.040	1819	4531
NVFP4	0.780 ± 0.041	2011	4103

AA-LCR (long-context reasoning) — 100 questions × 1 seed

config	acc (mean ± 95% CI)	median tokens	mean tokens
ThinkingCap-Qwen3.8-27B bf16	0.810 ± 0.077	844	1718
NVFP4	0.780 ± 0.081	920	1508

Decode speed and MTP self-speculative decoding (MMLU-Pro) — 32 questions × 1 seed, vLLM 0.26 on one H200, 16 concurrent requests

config	median tokens	tok/s	s / task	MTP speedup	accept_len (max 4)
Qwen3.8-27B base · standard	484	51.3	8.9	1.00×	—
Qwen3.8-27B base · MTP	502	91.0	4.2	1.77×	2.59
ThinkingCap-Qwen3.8-27B bf16 · standard	232	50.5	3.7	1.00×	—
ThinkingCap-Qwen3.8-27B bf16 · MTP	216	85.7	2.0	1.70×	2.60
NVFP4 · standard	216	61.0	2.6	1.00×	—
NVFP4 · MTP	215	74.9	1.7	1.23×	2.54

Where to find us

Need even more efficiency? The open release is production-ready. Our enterprise versions go further — fewer thinking tokens still, tuned to your workload, at matched accuracy on your own tasks. Built for AI labs, inference providers and enterprises running models at scale. Deployed on your infrastructure, or in the cloud and region you choose. Talk to our team