Qwen3.8-Flash-Next is the August 26, 2026 release that plays the role Qwen3-Next once played for Qwen3.5: an open-weight preview of the architecture the next generation — Qwen4 — will be built on. It is a multimodal mixture-of-experts with 125B main parameters, a separate 51B N-gram embedding table, and just 6B parameters active per token.
The headline claim is efficiency. Alibaba says training cost about one-ninth of Qwen3.7-Plus while coding and office-task scores went up, and that prefill throughput at a 1M-token context is 8.6x higher when the prefix cache hits 90% of the time. The production version, qwen3.8-flash, is priced at $0.15 / $0.47 per million tokens internationally.
What is new in the architecture
- GDN + QSA attention. Three of every four layers use Gated DeltaNet to compress history into a fixed state; the fourth keeps global attention, now replaced by Qwen Sparse Attention, whose lightweight indexer scores micro-blocks rather than single tokens. At 1M tokens, the QSA kernel is up to 7.6x faster in prefill and 4.9x in decode.
- Gated Residual. The residual stream is split into four dynamically gated branches, which Alibaba says improves cross-layer information flow and training stability; the residual state can be stored in FP8.
- N-gram embedding. A 51B-parameter lookup table keyed on short local contexts adds capacity with almost no per-token compute and can live in host memory with asynchronous prefetch.
- Muon optimizer and a re-fitted scaling law that tolerates larger learning rates and batches, so batch-size warmup was dropped.
Benchmarks
Vendor-reported by Alibaba (August 26, 2026), text-only table. Harnesses differ by row (Claude Code or mini-SWE-agent; the better score is reported for DeepSWE).
| Benchmark | Flash-Next (6B active) | Qwen3.8-27B | Qwen3.7-Plus | DeepSeek V4-Flash-0731 | Claude Opus 4.6 (Max) |
|---|---|---|---|---|---|
| SWE-bench Pro | 62.5 | 61.7 | 55.8 | 56.0 | 53.4 |
| DeepSWE v1.1 | 58.7 | 42.2 | 16.5 | 54.4 | — |
| SWE-bench Multilingual | 81.0 | 73.8 | 75.8 | — | 77.5 |
| NL2Repo-Bench | 48.1 | 42.3 | 41.1 | 54.2 | 47.6 |
| CoWorkBench | 73.9 | 70.7 | 65.1 | 45.1 | 68.2 |
| JobBench | 55.7 | 33.4 | 27.6 | 41.3 | 36.6 |
| Toolathlon Verified | 73.5 | 67.1 | 50.6 | 70.3 | — |
| IFBench | 81.3 | 79.5 | 79.1 | 79.2 | 62.5 |
| GPQA Diamond | 91.7 | 89.2 | 90.3 | 90.8 | 91.3 |
| HLE | 35.9 | 30.8 | 34.7 | 33.8 | 40.0 |
| LiveCodeBench v6 | 91.9 | 90.3 | 89.6 | 90.6 | 88.8 |
On the multimodal side Alibaba reports AndroidWorld 84.5, OSWorld 2.0 19.4 (binary) / 52.3 (partial), Vision2Web 64.0, RealWorldQA 88.5 and LVBench 76.6.
Pricing
| Per 1M tokens | International | Mainland China |
|---|---|---|
| Input | $0.15 | ¥0.8 |
| Output | $0.47 | ¥2.7 |
The open weights themselves are free to download under the Qwen Community License. Global-region and batch calls are discounted on Alibaba Cloud’s pricing page.
Specs
| Spec | Value |
|---|---|
| Parameters | 125B main + 51B N-gram embedding; 6B active per token |
| Architecture | MoE, GDN + Qwen Sparse Attention, Gated Residual, N-gram embedding, MTP |
| Context window | 262,144 native; up to 1M with YaRN (hosted qwen3.8-flash: 1M) |
| Modalities | Text, image and video in; text out |
| API model | qwen3.8-flash (reasoning effort xhigh / medium / low) |
| Weights | Qwen/Qwen3.8-Flash-Next and an FP8 build on Hugging Face and ModelScope |
| License | Qwen Community License 1.0 |
How it compares
- vs Qwen3.8-Max: Max is the 2.4T flagship; Flash-Next is the cheap, efficient architecture preview.
- vs GLM-5.3-Flash: a smaller 6B-active model at a similar price; GLM-5.3-Flash reports 63.4 on DeepSWE v1.1 against Qwen’s 58.7 (different harnesses) and is MIT-licensed.
- vs DeepSeek V4.1-Flash: DeepSeek’s newer 552B model is much larger and scores higher on DeepSWE v1.1 (74.2); Flash-Next is the lighter architecture play.
- vs Qwen3.8-Omni-Flash: Omni-Flash adds audio and focuses on audio-video agents.
Limitations
- Preview status. Alibaba says the Qwen4 family is yet to come.
- Older baselines. The table compares against Claude Opus 4.6, Qwen3.7-Plus and DeepSeek V4-Flash-0731, not the latest frontier models.
- Vendor-reported scores; harness settings vary by row.
- Community license rather than a standard open-source license.
For the wider picture, see the 2026 LLM guide.
Best for
- High-concurrency coding and office agents at low cost
- Tool-driven workflows with long context
- Multimodal agents for mobile and computer use (AndroidWorld 84.5)
- Research on the Qwen4 architecture and its sparse attention
Pros & cons
Strengths
- Only 6B parameters active per token yet SWE-bench Pro 62.5 and DeepSWE v1.1 58.7 in Alibaba's table — ahead of Qwen3.8-27B and the 13B-active DeepSeek V4-Flash-0731
- A preview of the Qwen4 architecture — Gated DeltaNet plus Qwen Sparse Attention, four-branch Gated Residual and a 51B N-gram embedding stored in host memory
- Trained at roughly 1/9 the cost of Qwen3.7-Plus, and 8.6x its prefill throughput at 1M context with a 90% prefix-cache hit
- Open weights on Hugging Face and ModelScope, an FP8 build, and a production API at $0.15 / $0.47 that plugs into Claude Code, Codex and Qwen Code
Limitations
- It is an architecture preview — Alibaba says the full Qwen4 family comes later, so treat this as a research and efficiency release rather than a flagship
- The Qwen Community License is not MIT or Apache — read it before commercial use
- The 125B main weights plus 51B N-gram table still need a multi-GPU server (or host-memory offload) to self-host
- Behind the frontier on hard knowledge — HLE 35.9 against Claude Opus 4.6's 40.0
Frequently asked questions
When was Qwen3.8-Flash-Next released?
Alibaba released the open weights on August 26, 2026 on Hugging Face and ModelScope, with a technical report on GitHub. The production version is served as `qwen3.8-flash` on the Qwen Cloud platform.
Is Qwen3.8-Flash-Next the same as Qwen3.8-Flash?
Almost. Flash-Next is the open-weight model; Qwen3.8-Flash is the production version that defaults to a 1M context and ships with built-in official tools. The API price is $0.15 per 1M input tokens and $0.47 output internationally.
What does "Next" mean here?
As with Qwen3-Next for Qwen3.5, it previews the architecture that the next generation will use. Flash-Next pairs Gated DeltaNet with Qwen Sparse Attention (QSA), expands the residual stream to four gated branches, adds N-gram embeddings and is trained with the Muon optimizer.
How big is it?
125B main parameters plus a 51B N-gram embedding table, with 6B parameters active per token. The embedding table uses deterministic lookups, so it does not enter the per-token matrix-multiply budget and can sit in host memory.
What context length does it support?
The open weights support 262,144 tokens natively and up to 1,000,000 with YaRN. The hosted qwen3.8-flash defaults to 1M.