Alibaba releases Qwen3.8-Flash, an open-weight, 125B-parameter model built on its next-gen Qwen 4 architecture, saying it rivals Opus 4.6 and V4-Flash
Alibaba Group Holding Ltd. released the latest model under its popular Qwen series, a lower-priced platform aimed at driving adoption of its marquee AI offering globally.
BloombergLuz Ding
Context & Ripple Effects
This is the third step in a rapid cadence: after debuting Qwen3.5 in February 2026 and shipping Qwen3.8 weights under Apache 2.0 less than two weeks ago, Alibaba is now using an open-weight release to preview its next architecture rather than to close out the current one. Qwen3.8-Flash-Next is a sparse multimodal mixture-of-experts — roughly 125B total parameters with a small active slice — built on what the team calls the Qwen 4 foundation: hybrid GDN-plus-QSA attention, gated residuals, n-gram embeddings, and a Muon-style optimizer.
The framing has shifted from capability alone to cost-efficiency: Alibaba's own posts tout a QSA attention kernel up to 7.6× faster in prefill at 1M-token contexts, and a production API tier planned at about $0.16 per million input tokens. Community reaction, including @kimmonismus's claim that the 6B-active model beat Claude Opus 4.6 Max on 8 of 9 comparable benchmarks, treats this as the first credible open-weight shot across a frontier lab's bow — though those are benchmark claims by partisans, not independently verified results.
First-order effects
Developers gain an Apache-2.0-licensed multimodal model whose planned QwenCloud price (~$0.16/1M input tokens) undercuts the prevailing cost of frontier closed models like Opus 4.6, immediately usable for self-hosting or cheap fine-tuning.
Anthropic and other closed-frontier labs face their first widely available rival claiming near-parity on coding-heavy benchmarks at a fraction of the serving cost.
Second-order effects
Rival labs are pushed to compete on inference economics — sparsity, long-context kernels, prefix caching — rather than raw benchmark scores, since Alibaba's pitch centers on throughput and price.
Cloud providers hosting open weights capture the demand that would otherwise flow to proprietary APIs, pressuring closed labs' enterprise distribution channels.
Third-order effects
If open weights keep arriving as architecture previews ahead of the flagship generation, the moat around closed frontier models narrows to months of latency advantage plus hosted convenience — a structural shift toward open-weight commercialization where free weights are the funnel and paid inference is the business.
Benchmark credibility becomes a battleground: vendor-published comparisons against named rivals will draw increasing scrutiny, pushing independent evaluation into a gating role for adoption decisions.
The trend: Open-weight releases are becoming the top of the commercial AI funnel — Alibaba gives away next-generation weights to seed adoption, then monetizes through a low-priced hosted API.
Qwen 3.8 Flash-Next official released: A 6B-active open model just beat Claude Opus 4.6 Max across 8 of 9 comparable benchmarks! Qwen3.8-Flash-Next is a highly sparse MoE: • 125B model parameters • 51B additional n-gram embeddings • Only 6B parameters active per token It
wait, Qwen just open-weighted an early preview of Qwen4!!! Qwen3.8 Flash Next with a next gen architecture! > GDN + QSA hybrid attention > Gated Residual > N-gram Embedding > Muon optimizer > used in Qwen4 running this now! let me know if you want the endpoint!
At a 1M-token context length, QSA's attention kernel is up to 7.6× faster in prefill and 4.9× faster in decode. With a 90% prefix-cache hit rate, Qwen3.8-Flash-Next delivers 8.6× the prefill throughput of Qwen3.7-Plus.
Model Architecture Four core upgrades for maximum capability, efficiency, capacity, and stability: - Attention: GDN + QSA Hybrid. Gated DeltaNet (GDN) compresses history. Qwen Sparse Attention (QSA) uses a lightweight indexer for micro-block context selection. Lower the cost of
⚡Meet Qwen3.8-Flash, a multimodal MoE and an early preview of the Qwen4 architecture, now open-weight! The production version Qwen3.8-Flash will be available soon via QwenCloud API at just $ 0.16/1M input tokens and $ 0.47/1M output tokens. 125B parameters + 51B N-gram
Qwen3.8-Flash-Next is here! An open-weight multimodal MoE built on a brand new architecture, with native 256K context extendable to 1M via YaRN. 🤖 https://www.modelscope.ai/... 🏆 Leads every compared model on SWE-bench Pro (62.5 vs 53.4 for Claude-Opus-4.6 Max), SWE-bench
So yeah, today is gonna be a big day for Chinese open source: - Ox Alpha will be officially released as an open weight model And - Qwen3.8-Flash-Next (~125B-A6B + 51B n-gram) Will be released. Good day for open source. Quo vadis, western AI? Do you want to respond to
Qwen3.8-Flash can now be run locally! 🔥 The 125B MoE model outperforms Claude-Opus-4.6 (Max). Run on 75GB RAM via Unsloth GGUFs. Qwen3.8-Flash-Next enables CPU RAM / unified mem setups to deliver near VRAM speeds. Guide: https://unsloth.ai/... GGUF: https://huggingface.co/...
Qwen3.8-Flash-Next ranks ~#8 in Code Arena: WebDev scoring 1617 (AutoEval) and ~#3 among open weights! This is an early score via AutoEval. We'll continue to track to see where it lands in rank as live votes come in.
Exciting news: Qwen3.8-Flash-Next by @Alibaba_Qwen ranks ~#8 in Code Arena: WebDev (#3 among open models) scoring 1617 (AutoEval). Priced at $0.16/$0.47 Mtokens, it reshapes the Code Arena: WebDev Pareto Frontier! This 125B parameter model (6B active) has been touted as a
The biggest differentiator for Qwen3.8-Flash-Next isn't just the performance gains but also the memory and bandwidth footprint. Being able to take a 1x DSV4 hardware setup and run 2x 3.8-Flash-Next with it at half the activated parameters per token is a genuine force multiplier.