bottlecapai/ThinkingCap-Qwen3.8-Flash-Next
Image-Text-to-Text âĒ 180B âĒ Updated âĒ 710 âĒ 41
ThinkingCap Qwen 3.8 Flash-Next
ThinkingCap Qwen3.8-Flash-Next: bf16, FP8 and NVFP4 for vLLM; GGUF and a mixed-precision RCO GGUF for llama.cpp.
Note ð bf16 reference, vLLM/Transformers, 360 GB; served on 4Ã H200 (tensor parallel 4). Against the base Qwen3.8-Flash-Next: 38.5% fewer reasoning tokens on average (22% to 51% by benchmark) at 88.8% average accuracy vs 89.8%. The baseline for the quantized builds below.
bottlecapai/ThinkingCap-Qwen3.8-Flash-Next-FP8
Image-Text-to-Text âĒ 180B âĒ Updated âĒ 232Note ðą FP8 in Qwen's own FP8 layout (routed experts and n-gram table), vLLM, 186 GB; served on 2Ã H200 (tensor parallel 2). Evals vs bf16: AIME 2026 +0.73 pp, RealWorldQA â0.69 pp, thinking +3.3% / â0.1%, none significant. MTP accepts 2.77 tokens per step, as bf16 does.
bottlecapai/ThinkingCap-Qwen3.8-Flash-Next-NVFP4
Image-Text-to-Text âĒ 180B âĒ Updated âĒ 1.2k âĒ 8Note ð NVFP4 weight-only routed experts, FP8 n-gram table, vLLM, 132 GB; served on 2Ã H200 or 2Ã RTX PRO 6000 (tensor parallel 2, Marlin kernel). Evals vs bf16: AIME 2026 +0.73 pp, RealWorldQA +0.56 pp, both within noise; thinks 18% longer on AIME 2026.
bottlecapai/ThinkingCap-Qwen3.8-Flash-Next-GGUF
Image-Text-to-Text âĒ 180B âĒ Updated âĒ 5.18k âĒ 21Note ðĶ GGUF with vision projector and MTP head, llama.cpp/LM Studio/Ollama: ðĶ IQ4_XS 95.5 GB, ð Q4_K_M 113.2, ð Q6_K 171.9, ð Q8_0 191.0, ðŧ BF16 359.2. llama.cpp keeps the n-gram table (29 GB in IQ4_XS) in system RAM. Evals vs bf16 (AIME 2026, RealWorldQA): no accuracy difference; quants below Q8_0 think about 10% longer on AIME. For parallel image requests use -np N --no-kv-unified.
bottlecapai/ThinkingCap-Qwen3.8-Flash-Next-GGUF-RCO-3.8bpw
Image-Text-to-Text âĒ 180B âĒ Updated âĒ 383 âĒ 2Note ðŠķ GGUF with per-tensor mixed precision (IST-DASLab's RCO allocation), 3.82 bits per weight, 85.6 GB (+0.9 GB mmproj); llama.cpp. Its weights take 56.5 GB of GPU memory, and one 96 GB RTX PRO 6000 holds the full 262k context with vision and MTP (64 GiB). Evals vs bf16: AIME 2026 â0.94 pp, RealWorldQA â0.16 pp, LiveCodeBench v6 +0.29 pp, none significant; thinks 14% longer on AIME, not on LiveCodeBench.