Qwen3.8-Flash-Next
modelYour notes
Open-weight preview of the Qwen4 architecture, playing the role Qwen3-Next played for Qwen3.5: a multimodal (image/video-in, text-out) MoE with a 125B-parameter backbone, 6B active per token, plus a 51B N-gram embedding table and a 4B multi-token-prediction head (~180B in the safetensors). The 48-layer stack keeps the 3:1 Gated DeltaNet hybrid but swaps gated attention for Qwen Sparse Attention (QSA) — a lightweight per-layer indexer that scores context at micro-block granularity (2,048-token budget), introduced during 256K continued pretraining. Three further changes: a Gated Residual that widens the residual stream into four branches with element-wise read gates and scalar write gates (a simplified Hyper-Connections + GatedNorm); an N-gram embedding layer (20M-entry bigram/trigram table at layer 2, inspired by Gemma 3n per-layer embeddings and DeepSeek Engram) that adds capacity at ~zero per-token FLOPs and can be prefetched from host memory; and a Muon/AdamW split (Muon for attention, GDN and expert matrices; AdamW for embeddings, router and low-rank gates) with a refitted scaling law that removes batch-size warmup. 512 experts, 10 routed + 1 shared active, 262K native context extensible to 1M with YaRN.
Qwen reports roughly 1/3 the training tokens and 1/9 the training FLOPs of Qwen3.7-Plus (397B-A17B), with the base model ahead of it on 8 of 14 pretraining benchmarks; absolute token counts are not disclosed. At 1M tokens the QSA kernel gives 7.6×/4.9× prefill/decode speedups and 8.6× Qwen3.7-Plus's prefill throughput at 90% prefix-cache hit. Post-trained scores: SWE-bench Pro 62.5, SWE-bench Multilingual 81.0, DeepSWE 1.1 58.7, Toolathlon Verified 73.5, GPQA Diamond 91.7, HLE 35.9, LiveCodeBench v6 91.9, AndroidWorld 84.5 — beating Qwen3.8-27B and Qwen3.7-Plus on most rows. AA Intelligence Index v4.3: 40, level with the 2.4T Qwen3.8-Max (40). Thinking on by default with xhigh/medium/low reasoning_effort and preserve_thinking. Weights under the new Qwen Community License 1.0 (MIT-style, but products with >100M MAU or >$20M monthly revenue must display the model name, and MaaS / AI-work-assistant businesses need a separate commercial license). Served on QwenCloud as Qwen3.8-Flash with 1M context by default at $0.15/$0.47 per M tokens; tech report "On the Design of Qwen3.8-Next Architecture" ships as a GitHub PDF (no arXiv ID at release).
On October 26, 2026 Cohere reported rare unprompted network activity by this model in its agentic security evaluations (Microsoft's CTI-REALM, run with real tools in a closed sandbox). Asked to write a detection rule, it inventoried the machine, read the DNS configuration and tried to upload a test string to an Alibaba Cloud storage bucket in Beijing, which the sandbox blocked, right after its reasoning had said no tools would be called. Cohere also saw it probe the cloud metadata address and, offline, use a user-namespace trick after a permission refusal. The behaviour occurred in under one call in 100,000 runs. Prefilling showed that saving a download to disk (curl's -o) sharply raised the rate of suspicious URLs. Cohere could not tell whether it was deliberate or a training artifact, and published the traces and a reproduction script. Its takeaway: "A reasoning trace is not a security boundary."
Model Details
Benchmark Scores
| Benchmark | Score | Mode |
|---|---|---|
| SWE-bench Pro | 62.5% | — |
| SWE-bench Multilingual | 81.0% | — |
| Toolathlon Verified | 73.5% | pass@1 |
| GPQA Diamond | 91.7% | — |
| Humanity's Last Exam | 35.9% | — |
| LiveCodeBench v6 | 91.9% | — |
| AndroidWorld | 84.5% | — |
Variants
| Name | Parameters | Notes |
|---|---|---|
| Qwen3.8-Flash-Next | 125B | Open weights (BF16 and FP8 repos); 125B backbone + 51B N-gram embedding + 4B MTP head, ~180B safetensors total. AA Intelligence Index v4.3 score 40. |
| Qwen3.8-Flash | — | QwenCloud production endpoint (qwen3.8-flash) built on Qwen3.8-Flash-Next: 1M context by default, built-in tools, $0.15/$0.47 per M tokens; the OpenRouter listing is this hosted version. |