bottlecapai/ThinkingCap-Qwen3.8-27B
Image-Text-to-Text β’ 28B β’ Updated β’ 609 β’ 110
ThinkingCap Qwen 3.8
Efficient ThinkingCap Qwen3.8, bf16 and FP8 for datacenters. Request access and contact us for enterprise use if happy.
Note π bf16 reference, vLLM/SGLang/Transformers, 56 GB. Runs on any bf16 GPU (Ampere or newer). Needs an 80 GB GPU (H100, A100): weights plus ~2 GB per 32k tokens of context. The accuracy and reasoning-length baseline for FP8 below and for the quantized builds (NVFP4, GGUF, MLX): https://huggingface.co/collections/bottlecapai/thinkingcap-qwen-38-quants-6ab534cf5aee57103f21390d
bottlecapai/ThinkingCap-Qwen3.8-27B-FP8
Image-Text-to-Text β’ 28B β’ Updated β’ 339 β’ 10Note π± FP8 block-wise, vLLM, 31.2 GB. Native on Hopper and Blackwell. Needs a 40 GB+ GPU (weights plus ~2 GB per 32k tokens of context). Evals on H200: accuracy β0.3 to +1.0 pp per benchmark, mean tokens β12% to +13%.