Qwen3.8-Flash-Next

// Overview

Qwen3.8-Flash-Next is the open-weight model Alibaba's Qwen team published on August 26, 2026. Qwen frames it not as a product tier but as an architecture release: the model card calls it "this experimental preview of the architecture that will underpin Qwen4", and it is the first open-weight model shipped under that design. It is a causal language model with a vision encoder, trained through both pre-training and post-training, and it operates in thinking mode by default.

The sparsity is the point. The model carries 125B parameters but activates only 6B per token, plus a 51B n-gram embedding table and a 4B multi-token-prediction head. In Qwen's own comparison it sits against Qwen3.8-27B (27B dense), Qwen3.7-Plus (397B total / 17B active) and DeepSeek-V4-Flash-0731 (284B total / 13B active) — the smallest activated footprint in the group — while leading most rows of that table.

Four architectural changes carry the release. The Gated DeltaNet and Gated Attention pairing used in Qwen3-Next is reworked into Gated DeltaNet plus Qwen Sparse Attention (QSA), which selects at the micro-block level rather than the individual token and so cuts long-context latency. Gated Residual modulates widened residual streams through an element-wise read gate and a per-branch scalar write gate. N-gram Embedding scales parameters along an axis that is cheap to compute and easy to offload. And the training recipe applies Muon and AdamW to specific weight categories while dropping batch-size warmup entirely, starting directly at the target batch size.

On the language side Qwen reports 62.5 on SWE-bench Pro, 81.0 on SWE-bench Multilingual, 58.7 on DeepSWE 1.1, 73.5 on Toolathlon Verified (pass@1), 91.7 on GPQA Diamond and 91.9 on LiveCodeBench v6, with 35.9 on HLE — the one headline row where Qwen's table puts Claude-Opus-4.6 (Max) ahead, at 40.0. Qwen notes that SWE-bench Pro figures other than the Claude column were re-evaluated in-house on a corrected version of the benchmark, and that its HLE runs are judged by GPT-4o.

The vision-language results are aimed at agent work rather than captioning: 84.5 on AndroidWorld for mobile use, 64.0 on Vision2Web for visual web development, 76.6 on LVBench for long-video understanding, 88.5 on RealWorldQA and 72.3 on ERQA for embodied reasoning. Two of Qwen's rows — RecreationBench for long-horizon application recreation and CoWorkBench for office and productivity agents — are in-house benchmarks it describes but does not publish.

Context length is 262,144 tokens natively, extensible to 1,000,000. The weights are on Hugging Face under the Qwen Community License 1.0, in Transformers format and an FP8 variant, with serving recipes documented for SGLang, vLLM and TokenSpeed. Qwen points production users at a separate hosted model, Qwen3.8-Flash, which it describes as the official version built on Qwen3.8-Flash-Next with a 1M context window by default and built-in tools.

// Benchmarks

This model's scores

1. SWE-bench Pro62.5%
2. SWE-bench Multilingual81%
3. DeepSWE 1.158.7%
4. Toolathlon Verified (pass@1)73.5%
5. GPQA Diamond91.7%
6. LiveCodeBench v691.9%
7. HLE35.9%
8. AndroidWorld (mobile use)84.5%
9. LVBench (long video understanding)76.6%

Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.

// Strengths

• Only 6B activated parameters per token — the smallest activated footprint in Qwen's own comparison table, against 13B–27B for the models it is set beside
• SWE-bench Multilingual of 81.0, ahead of the 77.5 Qwen lists for Claude-Opus-4.6 (Max)
• 62.5 on SWE-bench Pro and 58.7 on DeepSWE 1.1, both leading Qwen's published comparison
• Strong GUI and device agent results: 84.5 on AndroidWorld and 64.0 on Vision2Web
• 262K native context extensible to 1M, with Qwen Sparse Attention selecting at micro-block level to hold down long-context latency
• Open weights under the Qwen Community License 1.0, with SGLang, vLLM and TokenSpeed recipes

// Best for

• Agentic software engineering where per-token cost and long-context latency dominate the bill
• Multimodal GUI and device agents — mobile automation, computer use, visual web development
• Long-video and document understanding inside an agent loop rather than as a separate vision call
• Self-hosted deployment on memory-constrained accelerators, where the n-gram embedding table can be offloaded
• Evaluating the Qwen4 architecture ahead of the production line built on it

// Qwen-Next (architecture preview) — every version

The full lineage of the Qwen-Next (architecture preview) line, newest first. Every version has its own page — click any to compare specs, benchmarks and pricing.