Qwen3.8-Flash-Next

AA coding index 73.1 (Artificial Analysis, frozen since 2026-09-11)

Qwen's preview of the Qwen4 architecture (Qwen Community License, released 2026-08-24). 180B on paper, but that splits into 125B of model with 6B active, 51B of n-gram embedding and 4B of MTP. Attention is hybrid: only 12 of its 48 layers are classic attention, the other 36 are Gated DeltaNet, so the KV cache grows far slower with context than the layer count suggests. It also has a vision encoder.

BigCodeBench-Hard pass@1 32% (48/148), via official protocol · llama.cpp UD-IQ3_XXS · one request at a time · a shared cluster GPU · mean of 3 runs, range 0.0 pp. 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.

Each quant links to the exact file it was measured from, in unsloth/Qwen3.8-Flash-Next-GGUF. Every point measured on the same class of card, one pass, one request at a time. One problem is 0.68 points, so anything under ~3.4 points apart is a tie — that is the spread we measure between GPU classes, and our rows are not all on the same card. Each card column is the context left for the KV cache after the weights, KV-Q4 / KV-Q8. A dash means the file does not fit that card, or fits with no room left to work. Context is usually the real constraint, not quality. Computed for one slot (--parallel 1), which is what you get serving yourself. llama-server opens four by default, and on a hybrid that costs real context — the table below each curve shows how much. The card figures assume a dedicated card. Measured on an idle one: the driver keeps 623 MiB of 143,771, so 99.6% is usable. Compute buffers are measured too — a fixed ~130 MiB plus 1 MiB per 1K of context, the same slope on every architecture we checked. If your card also runs your desktop you get noticeably less, and we have not measured that case yet.

83.0% ± 2.4 over 1390 problems

UD-Q4_K_XL · a shared cluster GPU · KV f16 · 1 run. This is the non-agentic half of BFCL: does the model call the right function with the right arguments. It says nothing about how it behaves over a long agentic conversation, which is a separate measurement. The score is BFCL’s own: the unweighted mean of four groups (simple, multiple, parallel, parallel multiple), where simple is itself the mean of Python, Java and JavaScript. Every group counts the same whatever its size, and irrelevance detection — which we also measure, another 240 problems — is not part of it. The margin follows that same weighting.

Pick by the curve below, not by size: the biggest quant is not the best one here. UD-Q3_K_XL (83.8 GiB) beats UD-Q4_K_XL (103.7 GiB) outright, and the smallest of all, UD-IQ1_S, beats two files larger than itself. Two practical warnings from measuring it. It needs a build of llama.cpp with the qwen4_exp architecture (PR 27742 at the time of writing); the stock binary dies with `unknown model architecture`. And it crashes on a quantised KV cache — serve it with `-ctk f16 -ctv f16`, which also means it needs more memory for context than its size suggests. On the upside, attention is hybrid, so that KV grows about four times slower than the layer count implies. The 87 GiB UD-IQ4_XS runs on a 48 GB card by serving the experts from system RAM (`--n-cpu-moe`), which we measured does not change a single token of the output.