Qwen3.8-27B

Not independently scored by Artificial Analysis

A dense 27B (Apache-2.0) and a native vision-language model, released 2026-08-14. Dense means every parameter is read per token, so on a 16 GB card only the 3-bit Q3_K_M (~14 GB) fits fully; Q4 (17 GB) spills to CPU and halves the speed. It is a reasoning model, but thinking traces are huge (recommended budget up to 262K tokens), so at home you run it non-thinking.

BigCodeBench-Hard pass@1 27% (33/121), via Q3_K_M GGUF, non-thinking, on an RTX 4060 Ti (local). Full precision on a rented card scores higher — see the quant curve in tips.. The brutal counterpart to HumanEval — where HumanEval saturates near the top, BCB-Hard spreads the field, so this is the number that actually separates coding ability.

~36 tok/s on RTX 4060 Ti, measured by us. Q3_K_M + MTP speculative decoding via llama-server (--spec-type draft-mtp, ~87% draft acceptance); 18.3 tok/s without spec-decode. Ollama cannot do spec-decode.

Pull the GGUF (unsloth/Qwen3.8-27B-GGUF) and serve with llama-server, NOT Ollama, to get speculative decoding: llama-server -m Qwen3.8-27B-Q3_K_M.gguf --n-gpu-layers 99 --spec-type draft-mtp --spec-draft-n-max 2 --flash-attn on --jinja. The MTP draft head ships inside the GGUF and roughly doubles single-user speed.

Quant vs quality on our BCB-Hard (non-thinking, n=121): BF16 37.2%, FP8 43.0%, NVFP4 34.7% (all on a rented Blackwell), Q3_K_M 27.3% (local 16 GB). The 4-to-16-bit formats cluster together (the gaps are sampling noise at temp 0.7); the real cliff is 3-bit, which costs ~10 points. A bigger card that fits Q4/Q6/FP8 buys back most of the quality. No local GPU? Featherless AI is the only API route so far (not on OpenRouter, CrofAI or opencode Zen yet), but mind the context: its flat $25/mo Premium (Chat) plan caps context near 32K, useless for real agentic coding; the full 262K needs the token-based agentic plan ($25+/mo, 8 concurrent).