Alibaba Releases Qwen3.8-Flash: 125B-Parameter MoE With Only 6B Active, Training Cost Cut to 1/9, Input at 1 CNY/Million Tokens

Forum topic · 小凯 · 2026-08-26

Summary

On August 26, 2026, Alibaba's Qwen team released and open-sourced Qwen3.8-Flash, a 125B-total-parameter Mixture-of-Experts model that activates only 6B parameters per token. The model features a native 262K-token context window (expandable to 1M via YaRN), native multimodality, and a training cost reportedly one-ninth of the previous Qwen3.7-Plus. Pricing is set at 1 CNY per million input tokens and 3 CNY per million output tokens—roughly one-third the price of DeepSeek-V4-Flash and about 3% of Claude Opus 4.6. Key architectural upgrades include a hybrid GDN (Gated DeltaNet) plus QSA (Qwen Sparse Attention) attention design delivering up to 7.6x prefill and 4.9x decode speedup at 1M-token contexts, a four-branch Gated Residual mechanism with FP8 storage, a 51B offloadable N-gram embedding, and the Muon optimizer with re-fitted scaling laws. Benchmark claims include leads over Opus 4.6 on SWE-bench Pro (+9.1), JobBench (+20), AndroidWorld (+22.5), MathVision (+25.1), and ERQA (+31.5). Alibaba also open-sourced Qwen3.8-Flash-Next, described as the architectural prototype for the Qwen4 family, and launched the Qianwen Office productivity suite. Weights are available on Hugging Face and ModelScope.

Alibaba Releases Qwen3.8-Flash: 125B-Parameter MoE at 1 CNY/Million Input Tokens

On the evening of August 26, 2026, Alibaba's Qwen team officially released and open-sourced Qwen3.8-Flash, a sparse Mixture-of-Experts (MoE) model with 125B total parameters and only 6B activated per token, a native 262,144-token context window (expandable to 1M with YaRN), and native multimodality. The company says training cost is 1/9 of the previous Qwen3.7-Plus, and pricing is 1 CNY per million input tokens / 3 CNY per million output tokens—as low as one-third of DeepSeek-V4-Flash's price and roughly 3% of Claude Opus 4.6.

A simultaneous release, Qwen3.8-Flash-Next, is described as the prototype of a brand-new architecture that will form the basis of the Qwen4 model family. By open-sourcing the structural changes early, Alibaba invites community validation before building out Qwen4.

Key points

Efficiency revolution over parameter race: total parameters grew to 125B, but per-token activation is just 6B — knowledge capacity grows while inference compute stays comparable to a small dense model.
Four core architectural upgrades:
GDN (Gated DeltaNet) + QSA (Qwen Sparse Attention) hybrid attention: GDN compresses long history into fixed-size states; QSA uses a lightweight Indexer to select important context at micro-block granularity. At 1M-token contexts, the QSA attention kernel achieves up to 7.6x prefill and 4.9x decode speedup.
Gated Residual (GR): the residual stream is expanded into 4 parallel branches with dynamic gating, and residual state supports FP8 storage, cutting memory traffic for long contexts.
51B N-gram Embedding: extra embedding parameters that can be offloaded to host memory with asynchronous prefetching, expanding capacity without occupying GPU VRAM.
Muon optimizer plus re-fitted scaling laws adapted to the new architecture.
Benchmark claims (efficiency and low price without capability loss):

| Benchmark | Result | |---|---| | SWE-bench Pro (agentic coding) | +9.1 over Opus 4.6 | | CoWorkBench | above DeepSeek-V4-Flash | | Toolathlon Verified | above DeepSeek-V4-Flash | | JobBench | ~+20 over Opus 4.6 | | AndroidWorld | +22.5 over Opus 4.6 | | MathVision | +25.1 over Opus 4.6 | | ERQA (embodied AI) | +31.5 lead |

The pretrained Base model reportedly exceeds the 3x-larger Qwen3.7-Plus Base on SuperGPQA, GSM8K, and SWE-bench Pretrain.

Pricing: entering the "one-yuan era"

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

Per the post, input prices for flagship Qwen-tier models fell ~97.5% over 18 months (from ~40 CNY/M in 2024 Q3 to 1 CNY/M now), which the author frames as a tipping point for consumer AI affordability — 1 million input tokens covers roughly 500 typical chat sessions.

Qianwen Office and the "Next" release

The newly launched Qianwen Office suite runs on Qwen3.8-Flash in standard mode, covering a claimed 95% of daily office tasks (document summarization/translation, data analysis, PPT drafting, email, meeting notes, spreadsheet formulas). Harder tasks can fall back to Qwen3.8-Max/Pro.
Qwen3.8-Flash-Next ships with open weights on Hugging Face and ModelScope, letting the community vet the architecture before the full Qwen4 family (expected 2026 H2–2027 H1).
As of August 2026, the Qwen3.8 series (Max, 27B, Flash) totals 2.4T parameters; Qwen models have surpassed 3 billion downloads with over 300,000 derivative models.

Context in the Chinese open-source landscape

The post situates Qwen3.8-Flash within a maturing Chinese open-source matrix — Qwen3.8, Zhipu GLM-5.3 (with GLM-5.3-Flash at 320B-A18B), and DeepSeek V4 — spanning price points from ~1 to tens of CNY per million tokens. Against closed US frontier models priced around 25–30 CNY/M input, it argues Chinese open models have opened a distinct "affordable price band" rather than merely undercutting.

One flagged risk: sustained price wars could squeeze R&D budgets. The author argues the sustainable path is exactly what Qwen3.8-Flash demonstrates — architectural innovation (sparse activation, attention optimization, training efficiency) that lowers per-inference cost while growing total capacity.

Outlook

Over the next 12 months, the post anticipates: the full Qwen4 family on the Next architecture, deeper native multimodality (video understanding/generation, 3D), and continued agentic differentiation.

> Note: All figures, benchmark results, and pricing above are as reported in the original forum post citing Alibaba's August 26, 2026 announcements and Chinese media coverage.