Qwen3.8-Flash-Next

Qwen3.8-Flash-Next shipped as an open-weight-only architecture preview, not a production release (Hugging Face repo metadata: created 2026-08-24, last modified 2026-08-26). Despite the "Qwen3.8" version label, Hugging Face's own architecture metadata lists it as "Qwen4ExpForConditionalGeneration" (model_type "qwen4_exp") — the card frames it as an experimental preview of the architecture that will underpin Qwen4, not a Qwen3.8-generation model in the strict sense. No hosted API price exists for this exact checkpoint as of 2026-08-26 — no input or output price is listed anywhere, and a live check found no DashScope or Model Studio listing. The $0.16/1M input and $0.47/1M output figures circulating online belong to a different, not-yet-shipped sibling product: the bare "Qwen3.8-Flash" (without "-Next"), which the HF model card itself describes as "the official version based on Qwen3.8-Flash-Next with more production features, e.g., 1M context length by default, official built-in tools," to be served via Qwen Cloud (confirmed by a Qwen/Alibaba_Qwen X post, which is also the source of that pricing). This page's 262,144-token context window is the -Next preview's NATIVE length; the card separately says the preview is "extensible up to 1,000,000 tokens" via RoPE-scaling techniques such as YaRN, which is a different claim from the sibling's "1M context length by default" — extensible-with-work versus shipped-as-standard. Neither figure should be read as the other's. Licensing is "qwen-community-1.0" (Qwen Community License 1.0) per Hugging Face's structured API — a custom license granting broad commercial-use rights, distinct from the OSI-approved Apache-2.0 template; the same API confirms the checkpoint is gated:false, disabled:false, across 131 safetensors shards. Those shards also resolve an apparent size contradiction: Alibaba's headline "125B total, 6B active" describes only the Mixture-of-Experts core, while the full on-disk checkpoint totals roughly 180B parameters (179,999,981,459 across all dtypes — 179,999,981,424 of them BF16, per Hugging Face's safetensors metadata) once a 51B-parameter N-gram embedding layer and a 4B-parameter multi-token-prediction head are added in — both figures are accurate, they just answer different questions.

All four benchmark scores on this page are self-reported by Alibaba (source: the Hugging Face model card's benchmark table), and none has independent confirmation as of 2026-08-26. Live checks that day found the model listed on none of eight independent boards: artificialanalysis.ai (whose own search for "flash-next" surfaces only DeepSeek V4 Flash Vision), vals.ai, arcprize.org, livebench.ai, matharena.ai, deepswe.datacurve.ai (updated that same day), toolathlon.xyz, and snorkel.ai's Agents' Last Exam model filter. LiveBench in particular cannot add it soon on structural grounds: its snapshot cadence is six months and its current live release is dated 2026-06-25, two months before this model existed. The vendor numbers themselves also carry methodology flags worth noting: the 35.9 HLE score (thinking mode, the model's default configuration) is, per the card's own footnote, "judged by GPT-4o" — a different judge than Artificial Analysis uses in its independent HLE runs; the 58.7 DeepSWE 1.1 figure is the higher of two harnesses (Claude Code vs. mini-SWE-agent), and the card notes the model does best specifically on mini-SWE-agent — the same harness the official DeepSWE board itself uses, which is why a slot between claude-opus-4.8[max] (59%±2%) and qwen3.8-max[xhigh] (57%±3%) is a reasonable placement if the board, last updated the same day but still without this model, ever adds it; and the 24.3 Agents' Last Exam figure is Pass@1, chosen here over the vendor's separate 51.2 composite "Score" because Snorkel AI's own board ranks primarily by Pass Rate. All four numbers should be read as vendor-claimed pending independent re-confirmation on this exact checkpoint, not as settled scores.