Model Card｜Qwen3.8-Flash-Next

🌏 中文版

Model Information

Key Capabilities

• First public release of the Qwen4-generation architecture: Qwen Sparse Attention (QSA), which operates at the micro-block level rather than selecting individual tokens, replaces Gated Attention in the previous Gated DeltaNet + Gated Attention pairing — Alibaba says this significantly cuts long-context latency
• Adds a Gated Residual mechanism that modulates widened residual streams via per-branch read/write gates, improving expressiveness across deep layers while preserving training stability with low inference overhead
• Introduces a 51B-parameter N-gram Embedding (indexed by bigrams/trigrams), shifting part of the parameter scaling onto an axis that's cheaper computationally than MoE and friendlier to offloading on memory-constrained hardware
• Alibaba states training cost is roughly one-ninth of Qwen3.7-Plus's, yet the model matches or beats the 397B-parameter Qwen3.7-Plus on multiple coding/agent benchmarks

Benchmark Performance

⚠️ All figures above are from Qwen's own official model card (some evaluated via third-party harnesses like Claude Code / mini-SWE-agent, but run and compiled by the Qwen team). No independent third-party reproduction or standardized leaderboard listing exists as of publication.

Comparison with Predecessors and Competitors

Against its own 397B-parameter Qwen3.7-Plus, Qwen3.8-Flash-Next needs only 6B active parameters to score far higher on agent-oriented benchmarks like DeepSWE 1.1 (58.7 vs. 16.5) and CoWorkBench (73.9 vs. 65.1). This isn't simple parameter stacking — it's the "architectural efficiency" path Alibaba emphasizes: for the same compute budget, QSA sparse attention and Gated Residual trade for better long-context utilization. Notably, it even beats its own 27B dense model, Qwen3.8-27B, on most metrics, suggesting the MoE + sparse attention combination isn't just cheaper — it may be a genuine capability gain for agent tasks.

Against closed-source frontier models, Qwen3.8-Flash-Next leads or matches on real-world tool use (Toolathlon: 73.5 vs. DeepSeek-V4-Flash-0731's 70.3, Claude-Opus-4.6 unlisted) and CoWorkBench (73.9 vs. Claude-Opus-4.6's 68.2), but still trails on more frontier, general-reasoning-heavy tasks like Agents' Last Exam (24.3 vs. DeepSeek's 25.2) — suggesting the architectural efficiency gains show up mainly in execution on known task types, not a wholesale leap in general reasoning.

Pricing is the biggest open question right now: as an open-weight architecture preview, Flash-Next itself has no published Qwen Cloud hosting price. The only published figure belongs to the "production counterpart," Qwen3.8-Flash (default 1M context, built-in tools), listed at ¥1/¥3 per million tokens (roughly $0.16/$0.47) — but Alibaba hasn't clarified whether the two share the exact same architecture and parameter count, so they shouldn't be treated as equivalent.

Implications for Agent Development

The QSA sparse attention plus Gated DeltaNet combination is clearly aimed at cutting latency and cost for long-context agent workloads — a real pain point for most agent frameworks today, where longer contexts compound latency and cost with every tool call. The N-gram Embedding shifts part of the model's knowledge capacity onto a computationally cheaper axis, which in theory gives smaller teams a shot at running near-frontier agent capability on limited hardware.

• If you're building long-horizon, multi-tool-call agents (e.g., office automation tasks like CoWorkBench): Qwen3.8-Flash-Next's official scores are currently the strongest in its weight class, worth adding to your self-hosted or private-deployment evaluation shortlist
• If you're doing local/private deployment and care about memory footprint: 6B active parameters plus an offload-friendly N-gram Embedding design should, in theory, fit more easily on constrained hardware than a comparable dense model — though the total BF16 weight size is still roughly 180B, so high-end hardware or quantization is still required in practice
• Not suitable for: production lines needing an immediately stable API service — right now there are only open weights and community endpoints, no official SLA, and Alibaba itself labels this "an experimental architecture preview," distinct from the production Qwen3.8-Flash service

Takeaway

I used to read "Qwen4 architecture preview" as marketing language, but this time the Qwen team shipped the full architectural detail (QSA, Gated Residual, the exact N-gram Embedding configuration) together with official benchmarks in one release — rarer than the usual "tease first, document later" pattern. It's also a verification lesson: Flash-Next (the open-weight architecture preview) and Qwen3.8-Flash (the Qwen Cloud–hosted production version) have near-identical names, but Alibaba hasn't confirmed they share the same architecture — so their pricing, SLA, and even capabilities can't be assumed interchangeable. When writing a model card, "same name, different thing" cases need to be handled separately rather than assumed equivalent just because the names look alike.

References

• Qwen3.8-Flash-Next Model Card — Hugging Face
• Unite.AI: Qwen3.8-Flash-Next Previews Qwen4 Architecture With 6B Active Parameters
• Startup Fortune: Alibaba's Qwen3.8-Flash-Next Gives Builders An Early Look At Qwen4
• byteiota: Qwen 3.8-Flash-Next: Inside the Qwen4 Architecture Preview
• OrcaRouter: Qwen3.8-Flash-Next Is Out — Qwen4 Architecture Confirmed (includes Qwen3.8-Flash production pricing)
• Baekpica/Qwen3.8-Flash-Next-GGUF (confirms Qwen Community License 1.0 terms)