Qwen announced Qwen3.8-Flash-Next, open weights described as "an early preview of the Qwen4 architecture," alongside a separate, not-yet-released production model called Qwen3.8-Flash that will be priced on the QwenCloud API at $0.16 per million input tokens and $0.47 per million output. Two documents back the open-weight release: a 28-page technical report covering the base model's architecture and pretraining, and a separate set of benchmark tables published with the model release covering the post-trained, multimodal checkpoint's downstream performance. They answer different questions, and it's worth counting each on its own terms rather than treating the tweet's summary as the full picture of either.
The pretraining result: real, and well-documented
The technical report — "On the Design of Qwen3.8-Flash-Next Architecture: Evaluation, Efficiency, and Training Stability" — covers four design choices in real depth: a token-mixing hybrid of Gated DeltaNet (GDN) and standard attention (one full-attention layer per four, replaced by Qwen Sparse Attention during continued pretraining), a "Gated Residual" that widens the residual stream to four branches read through an elementwise gate, an n-gram embedding layer whose tables are prefetched from host memory rather than sitting on the accelerator, and a refitted Muon-optimizer recipe. The report keeps three evaluation axes distinct throughout — loss and downstream benchmarks, training/inference cost, and training stability — and explicitly reports cases where they disagree (enlarging the n-gram vocabulary lowers loss monotonically while downstream accuracy saturates; a sparse-write variant looked nearly free in pretraining loss and degraded only later). That kind of transparency about where a lab's own metrics disagreed is worth crediting directly.
Counting Table 11 directly: Flash-Next-Base (125B total parameters, 6B activated, plus 51B parameters of n-gram embedding tables held off-accelerator) beats the smaller Qwen3.8-27B-Base on all 14 pretraining benchmarks, and against the much larger, 397B/17B-active Qwen3.7-Plus-Base, it leads on 8 and trails on the remaining 6 by at most 2.6 points — while using roughly a third of the activated parameters, a third of the training tokens, and about a ninth of the training FLOPs. Qwen Sparse Attention tracks full attention's training loss within roughly 1×10⁻⁴ while running 7.6x faster in prefill and 4.9x faster in decode at a 1M-token context, at the kernel level; the FlashQLA kernel library gives a 2–3x forward and roughly 2x backward speedup over the existing Triton baseline. That's a real, specific efficiency claim about the base model, backed by numbers rather than adjectives.
The post-trained model's benchmark tables: mostly true, with three specific exceptions
Separately, Qwen published benchmark tables comparing Qwen3.8-Flash-Next against Qwen3.8-27B, Qwen3.7-Plus, DeepSeek-V4-Flash-0731, and Claude-Opus-4.6 (Max) across coding, agent, general-reasoning, and multimodal categories — with real methodology footnotes attached to nearly every row: disclosed harnesses (Claude Code, mini-SWE-agent), disclosed temperature/top_p/context settings, a disclosed judge model per benchmark where relevant (GPT-4o for HLE, a dated Claude Code build for Vision2Web), an explicit note that "problematic tasks were corrected and all baseline models were re-evaluated on the refined benchmark" for SWE-bench Pro rather than comparing a new run against old baseline numbers, and a specific reward-hacking countermeasure for NL2Repo-Bench (disabling Bash commands that could fetch or clone the target repository directly). That's a genuinely more rigorous disclosure than most self-published benchmark tables get, and worth naming as such — with one caveat worth flagging plainly: two of the benchmarks (CoWorkBench, RecreationBench) are described as in-house, meaning Qwen built the test it's grading itself on, which deserves more skepticism than an externally maintained one.
Counting both tables directly — 22 benchmark rows in total (12 in the coding/agent/general table, 10 in the multimodal table) — Qwen3.8-Flash-Next posts the best or tied-best score in 20. The tweet's own framing was that Flash-Next outperforms Qwen3.7-Plus "across the board" — that's not quite what the tables show once every row is checked rather than the headline five. Two rows go fully to another model: DeepSeek-V4-Flash-0731 wins NL2Repo-Bench outright (54.2 vs. Flash-Next's 48.1), and Claude-Opus-4.6 (Max) wins HLE, the hardest general-reasoning benchmark in either table (40.0 vs. Flash-Next's 35.9). Two more rows split between sub-metrics: on Agents' Last Exam, DeepSeek edges Flash-Next on the Pass@1 figure (25.2 vs. 24.3) while Flash-Next leads the benchmark's separate Score metric (51.2, no DeepSeek score reported); and on CharXiv, Qwen3.7-Plus itself — the model Flash-Next is supposed to beat "across the board" — leads without chain-of-thought prompting (85.8 vs. 84.6), while Flash-Next retakes the lead once chain-of-thought is allowed (90.6 vs. 85.9). None of these exceptions overturn the overall picture — 20 of 22 is a strong result — but they're a specific, useful correction to a blanket claim, and they follow a pattern worth noting on its own: Flash-Next's losses cluster on the hardest pure-reasoning row (HLE) and on narrow technical sub-metrics, not on the broad agentic and coding categories where its lead is largest and most consistent.
Training stability, credited plainly
One more result worth naming from the technical report: Qwen states that full-scale training of Qwen3.8-Flash-Next proceeded "without a single loss spike or anomalous fluctuation in gradient norms," without needing explicit stabilization tricks like qk-clip or SwiGLU-clip that other labs (including Kimi's team, cited directly in the report) have used to control training instability. That claim is backed by an actual stress test — running at four times the optimal learning rate and showing the new recipe stays stable where the previous architecture spikes — rather than asserted on its own.
What to expect next
- Watch for independent evaluation of the released weights. Both the pretraining and post-training numbers are Qwen's own; a third party running its own suite on the downloadable checkpoint is what would confirm the 13-of-16 sweep and the specific exceptions hold up outside Qwen's own harness choices.
- Watch the in-house benchmarks specifically. CoWorkBench and RecreationBench are Qwen's own constructions; independent replication matters more for those two rows than for the externally maintained benchmarks in the same tables.
- Watch for the production Qwen3.8-Flash release and whether its scores match Flash-Next's. The two were announced together under confusingly similar names; whether the priced, closed version posts the same numbers as the open-weight preview is worth checking once it ships.
References: Qwen (@Alibaba_Qwen) on X — announcement · Technical report — On the Design of Qwen3.8-Flash-Next Architecture · GitHub — QwenLM/Qwen3.8-Flash-Next · Hugging Face — Qwen/Qwen3.8-Flash-Next · related coverage: Qwen3.8-27B's Open Weights · Qwen3.8-Max's Open Weights Finally Landed · Frontier Arcade: trends & predictions