Qwen3.8-Flash-Next: Open Weights, Qwen4 Architecture

Qwen posted weights, not a Cloud SKU. Qwen3.8-Flash-Next is an experimental preview of the architecture that will underpin Qwen4. 6B activated, a 51B n-gram table, 4B MTP, 262K native. The managed Qwen3.8-Flash API is a different product built on top of this, with 1M context by default and official tools. If you came here for a list price, there is not one yet.

The language table they published puts this next to DeepSeek-V4-Flash-0731 and their own Qwen3.8-27B / Qwen3.7-Plus. DeepSWE and SWE-Pro move. NL2Repo does not. Treat every number as self-reported.

At a Glance

• Catalog id: qwen3.8-flash-next
• HF: Qwen/Qwen3.8-Flash-Next
• Release: August 26, 2026
• License: qwen-community-1.0 (HF card)
• Params: ~180B stored; 6B active + 51B n-gram + 4B MTP (125B with 6B activated)
• Context: 262,144 native / 1M YaRN
• Modalities: text, image, video in; text out
• Thinking: on by default; effort xhigh / medium / low (xhigh default); preserve_thinking on by default
• Price: none published. Catalog-only. Not on OpenRouter / fal at writeup.
• Not the same id as Qwen Cloud Qwen3.8-Flash

What's New

Open-weight Qwen4-architecture preview. QSA + Gated Residual + n-gram embedding. The n-gram table is the unusual scaling lever: parameters that are cheaper to offload than experts.

Not the Cloud Flash API. Same family, different product. 1M-by-default and built-in tools live on the hosted SKU.

Self-reported coding/agent scores that beat Flash-0731 on DeepSWE and SWE-Pro, plus a full vision suite.

Benchmarks

All figures below are self-reported on the HF README (as of August 26, 2026). Not LLM Stats verified.

DeepSWE +4.3 and SWE-Pro +6.5 vs Flash-0731 are the reason this is a post. LiveCodeBench v6 91.9 vs 90.6 and GPQA Diamond 91.7 vs 90.8 are near-saturated; the deltas are noise on a crowded ceiling. NL2Repo 48.1 vs 54.2 is the miss; say it out loud. CoWorkBench 73.9 vs 45.1is in-house, so it stays off the main chart. Agents' Last Exam pass@1 24.3 vs 25.2 is a wash.

Harness notes from the README, kept short: DeepSWE uses Claude Code and mini-SWE-agent; they report the higher, and Flash-Next bests on mini-SWE-agent. SWE-Pro: they re-evaluated baselines on a refined set except Opus 4.6 Max (official). CoWorkBench and JobBench / RecreationBench are in-house. HLE judged by GPT-4o.

AndroidWorld 84.5 vs Opus-4.6 Max 62.0is the vendor's loudest vision line. OSWorld 2.0 binary 19.4 is still a coin-flip on full success (tied with Qwen3.8-27B). MathVision without CI is 90.6, near the 27B; with CI it is 95.7. RealWorldQA 88.5 vs Opus 73.9 rounds out the callouts.

Architecture

Causal LM with a vision encoder. Text + image + video in, text out. Hidden 2560, vocab 248320 padded, 48 layers. Layout: 12 × (3 × (Gated DeltaNet → MoE) → 1 × (Qwen Sparse Attention → MoE)). MoE: 512 experts, 10 routed + 1 shared.

Hybrid Attention with QSA: Gated DeltaNet paired with Qwen Sparse Attention at the micro-block level (budget 512 blocks / 2048 tokens), not token-by-token selection. Gated Residual widens residual streams with 4 branches. N-gram embedding at layer 2 indexes 20,000,000 bigrams/trigrams; that is the param scale that is cheaper to offload than MoE experts. MTP is 1 layer. Training recipe: Muon + AdamW on specific weight categories. We are summarizing the card and the cited tech report (On the Design of Qwen3.8-Next Architecture, Qwen Team, August 2026), not re-deriving it.

Serving

Thinks by default. Sampling: thinking temp 1.0 / top_p 0.95; instruct 0.7 / 0.8. YaRN to 1M only when you need it (static YaRN can hurt short prompts). Compatible with Transformers, vLLM, SGLang, TokenSpeed. Not deployed by any Inference Provider on HF at writeup. No hosted $ yet.

When to Use It

Good fit

• Self-host / vLLM shops that want a 6B-active multimodal agent with a 262K window
• Coding loops where DeepSWE / SWE-Pro matter
• Vision + tools without a Cloud bill

Prefer Cloud Qwen3.8-Flash

When you want 1M default and official tools and do not want to run ~180B of weights.

Prefer Flash-0731 API

When you already pay DeepSeek and NL2Repo is your job.

Do not treat experimental Next as the production Cloud id.

Outlook

This post ages when Qwen Cloud prices a Flash-Next endpoint, when LLM Stats verifies the coding suite, or when Qwen4 stops being a preview sentence. The architecture claim is the through-line.

Sources: HF card, Qwen blog, GitHub, tech report citation On the Design of Qwen3.8-Next Architecture (Qwen Team, August 2026).