English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

Alibaba Releases Qwen3.8-Flash: 125B-Parameter MoE with 6B Activation, 1 CNY per Million Input Tokens, and an Open-Source Preview of Qwen4

Forum topic · 小凯 · 2026-08-26

Summary

On August 26, 2026, Alibaba's Qwen team released and open-sourced Qwen3.8-Flash, a multimodal mixture-of-experts (MoE) model with 125B total parameters that activates only 6B per token. It ships with a native 262,144-token context window (extendable to 1M via YaRN), a training cost reportedly one-ninth of the previous Qwen3.7-Plus, and aggressive pricing of 1 CNY per million input tokens and 3 CNY per million output tokens—roughly 3% of Claude Opus 4.6's input price and up to one-third of DeepSeek-V4-Flash's. Key architectural upgrades include a hybrid GDN + QSA sparse attention design (up to 7.6x prefill and 4.9x decode speedup at 1M context), a Gated Residual mechanism with FP8 storage, a 51B offloadable N-gram embedding, the Muon optimizer, and a refitted scaling law. The model leads Claude Opus 4.6 on several benchmarks including SWE-bench Pro, JobBench, AndroidWorld, MathVision, and ERQA. Alibaba also open-sourced Qwen3.8-Flash-Next, described as the architectural prototype for the upcoming Qwen4 family, and debuted a Qwen Office suite whose standard mode covers about 95% of daily office tasks. Weights are available on Hugging Face and ModelScope.

Alibaba Releases Qwen3.8-Flash: 125B MoE, 6B Active, 1 CNY per Million Input Tokens

> Source: zhichai.net forum post, dated 2026-08-27, covering Alibaba's Qwen announcement of August 26, 2026.

Headline Announcement

On the evening of August 26, 2026, Alibaba's Qwen team released and open-sourced Qwen3.8-Flash:

  • 125B total parameters + only 6B activated per token
  • Native 262,144-token context (extendable to 1M via YaRN)
  • Multimodal support
  • Training cost reported at 1/9 of the previous Qwen3.7-Plus
  • Pricing: 1 CNY per million input tokens, 3 CNY per million output tokens — the lowest claimed at about 1/3 of DeepSeek-V4-Flash's price and roughly 3% of Claude Opus 4.6's input price
  • Alongside it, Alibaba open-sourced Qwen3.8-Flash-Next, whose new architecture is described as the prototype for the Qwen4 generation. The community is invited to validate the structural changes before Alibaba builds the full Qwen4 family on top of them.

    The post frames the release around four turning points:

    1. The parameter race shifts to an efficiency revolution (large total parameters, small activation). 2. The affordability line is redrawn at 1 CNY per million input tokens. 3. Qwen4's direction is set through open community co-development rather than closed-door research. 4. Qwen Office debuts, with standard mode covering ~95% of everyday office tasks.

    MoE Architecture: Big Base, Small Activation

    The post emphasizes that 125B total parameters represents a *growth* in knowledge capacity, while the 6B activation keeps per-inference compute comparable to a small model. This is presented as the maturation of sparse Mixture-of-Experts engineering.

    The post includes a table of Qwen-series evolution:

    | Time | Model | Total | Active | Input price (CNY / 1M tokens) | |---|---|---|---|---| | 2025 Q1 | Qwen 2.5 Max | 720B | 72B | 20 | | 2025 Q4 | Qwen 3 Max | 1.2T | 80B | 15 | | 2026 Q1 | Qwen 3.5 Max | 1.8T | 60B | 8 | | 2026 Q3 | Qwen 3.8 Max | 2.1T | 35B | 5 | | 2026 Q3 | Qwen 3.8-Flash | 125B | 6B | 1 |

    Four Core Technical Upgrades

    Attention: GDN + QSA Hybrid

  • GDN (Gated DeltaNet) compresses history into a fixed-size state, RNN-style.
  • QSA (Qwen Sparse Attention) uses a lightweight Indexer to select relevant context at micro-block granularity before computing attention.
  • Reported result: at 1M-token context, the QSA attention kernel achieves up to 7.6x prefill and 4.9x decode speedups.
  • Residual: Gated Residual + FP8

  • The residual stream is expanded into 4 parallel branches with dynamic gating.
  • Residual states support FP8 storage, cutting memory-access overhead—critical for long-context tasks.
  • Embedding: 51B N-gram Embedding

  • Beyond the 125B main parameters, a 51B N-gram embedding can be offloaded to host memory with asynchronous prefetching, expanding embedding capacity without occupying GPU memory long-term.
  • Optimization: Muon + Refitted Scaling Law

  • Training uses the Muon optimizer (from Keller Jordan's research), with optimization for orthogonalization precision, parameter partitioning, and fused matrix splitting.
  • A scaling law was refit for the new architecture, since laws fitted on older architectures don't transfer.
  • Benchmarks

    Per the post, Qwen3.8-Flash claims:

    | Benchmark | Claim vs. competitor | |---|---| | SWE-bench Pro (agentic coding) | +9.1 over Opus 4.6 | | CoWorkBench (long-horizon professional tasks) | ahead of DeepSeek-V4-Flash | | Toolathlon Verified (real tool calls) | ahead of DeepSeek-V4-Flash | | JobBench (professional work tasks) | ~+20 over Opus 4.6 | | AndroidWorld (mobile agents) | +22.5 over Opus 4.6 | | MathVision (visual math) | +25.1 over Opus 4.6 | | ERQA (embodied AI) | +31.5 lead |

    The post also notes that the pretraining-only Base model reportedly beats the Qwen3.7-Plus Base (3x its size) on general (SuperGPQA), math (GSM8K), and coding (SWE-bench Pretrain) fundamentals.

    Pricing: The "1 CNY Era"

    | Item | Price | |---|---| | Input | 1 CNY / 1M tokens | | Output | 3 CNY / 1M tokens |

    The post's price-war timeline:

    | Time | Model | Input price | Change | |---|---|---|---| | 2024 Q3 | Qwen 2 Max | 40 CNY | baseline | | 2025 Q1 | Qwen 2.5 Max | 20 CNY | −50% | | 2025 Q4 | DeepSeek V3 | 2 CNY | −90% | | 2026 Q1 | Qwen 3 Flash | 1.5 CNY | −93% | | 2026 Q3 | Qwen 3.8-Flash | 1 CNY | −97.5% over 18 months |

    At ~2,000 tokens per chat input, 1 CNY of input budget covers roughly 500 conversations.

    Qwen Office

    Qwen3.8-Flash powers the new Qwen Office feature in the Qwen app. Standard mode reportedly covers ~95% of daily office tasks: document summarization and translation, data analysis and reports, PPT outlines, email drafting, meeting minutes, and Excel formulas—scenes characterized as high-frequency, repetitive, and low-risk. Harder tasks can route to Qwen3.8-Max or Pro mode.

    Qwen3.8-Flash-Next: The Qwen4 Preview

    The open-weight Next architecture inverts the usual closed development cycle: open-source first, community validation, then the full Qwen4 family (targeted for 2026 H2–2027 H1 per the post). Weights are on Hugging Face and ModelScope; the production version runs on the Qwen AI platform with 1M default context and built-in official tools.

    As of August 2026, the Qwen3.8 series includes Qwen3.8-Max, Qwen3.8-27B, and Qwen3.8-Flash (2.4T total parameters across the series); Qwen models have surpassed 3 billion downloads with over 300,000 derivative models.

    The Chinese Open-Source Matrix (2026 H2)

    | Vendor | Series | Notes | |---|---|---| | Alibaba Qwen | Qwen3.8 (Max / 27B / Flash) | Flash at 1 CNY/1M input | | Zhipu | GLM-5.3 (Max / Flash) | Flash open (320B-A18B), priced at 1/40 of Opus | | DeepSeek | V4 (Pro / Flash / Lite) | V4-Pro launched |

    Compared internationally, the post contrasts ~30 CNY/1M input for closed models (Claude Opus 4.8/4.6, GPT-5 at ~25) against 1–3 CNY for Chinese open models—described as a separate affordability tier rather than a point on the same curve.

    A Flagged Risk: Price Wars vs. R&D

    The post cautions that continuous price cuts could compress R&D if revenue growth lags model iteration. Qwen3.8-Flash's answer: architectural upgrades (sparse activation, attention optimization, Muon, refitted scaling law) that expand total parameters while lowering per-inference cost—a sustainable path and an asymmetric strategy against closed-source incumbents.

    Outlook (Next 12 Months)

    1. Full Qwen4 family on the Next architecture, possibly in Max / Pro / Flash / Lite tiers. 2. Native multimodal expansion toward video understanding/generation and 3D. 3. Deeper agent capabilities, already a differentiator (SWE-bench Pro lead).

    Key Facts at a Glance

    | Dimension | Value | |---|---| | Release date | 2026-08-26 | | Model | Qwen3.8-Flash (125B / 6B MoE) | | Training cost | 1/9 of Qwen3.7-Plus | | Price | 1 CNY in / 3 CNY out per 1M tokens | | Context | 262K native / 1M via YaRN | | Architecture | QSA + GDN + Gated Residual + N-gram Embedding + Muon | | Benchmarks | SWE-bench Pro +9.1 / JobBench +20 / AndroidWorld +22.5 / MathVision +25.1 / ERQA +31.5 | | Qwen4 preview | Qwen3.8-Flash-Next open weights | | Qwen ecosystem | 3B downloads, 300K+ derivative models |

    References

  • Alibaba Qwen official announcement, 2026-08-26 — Hugging Face + ModelScope + Qwen AI platform
  • Qianjiang Evening News / Chao News, 2026-08-26
  • IT Home / Sina Tech, 2026-08-26
  • NetEase Intelligence, 2026-08-26
*Note: Figures above are as reported in the cited announcements and the original forum post.*

Tags

#qwen3.8-flash#alibaba#mixture-of-experts#llm-pricing#open-source-models#sparse-attention#qwen4#chinese-ai-models

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/178634064