Alibaba Releases Qwen3.8-Flash: Massive MoE Efficiency at 1 CNY per Million Tokens
On the evening of August 26, 2026, Alibaba's Qwen team officially released and open-sourced Qwen3.8-Flash: a 125B-parameter Mixture-of-Experts model that activates only 6B parameters per token, with a native 262,144-token context window (expandable to 1M via YaRN), native multimodality, and a training cost reported at just 1/9 of the previous Qwen3.7-Plus.
Pricing is set at 1 CNY per million input tokens and 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 companion release, Qwen3.8-Flash-Next, ships open weights with a next-generation architecture that Alibaba describes as the prototype for the upcoming Qwen4 model family. The strategy shifts Qwen4's development from closed-door research to community co-validation.
Four headline takeaways:
1. The parameter race shifts to an efficiency revolution: 125B total / 6B active shows that strong models need not be compute-hungry. 2. The price-performance line is redrawn: 1 CNY per million input tokens pushes "inclusive AI" to an unprecedented low. 3. Qwen4 is pre-defined by open source: the Next architecture invites community scrutiny before the full Qwen4 family is built. 4. Qianwen Office debuts: standard mode reportedly covers 95% of daily office tasks.
MoE Architecture: Large Base, Small Activation
The key point of "125B total + 6B active" is not parameter reduction—total parameters grew to 125B, but each inference activates only 6B. Knowledge capacity expands while per-inference compute stays comparable to a traditional 6B model. This is the 2026-era engineering realization of MoE: many expert networks, with only the most relevant experts activated per task.
Qwen series evolution cited in the post:
| Time | Model | Total Params | Active Params | Input Price (CNY/M 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
Standard attention scales at \(O(n^2)\). Qwen3.8-Flash combines:
- GDN (Gated DeltaNet): compresses long history into a fixed-size state via RNN-like state compression.
- QSA (Qwen Sparse Attention): a lightweight Indexer selects important context at micro-block granularity before full attention is computed.
- [1] Alibaba Qwen official announcement, "Qwen3.8-Flash released and open-sourced," 2026-08-26 — Hugging Face + ModelScope + Qwen AI platform
- [2] Qianjiang Evening News / Chao News, 2026-08-26 — https://www.toutiao.com/article/7678325863821230626/
- [3] IT Home / Sina Tech, 2026-08-26 — https://tech.sina.cn/2026-08-26/detail-inipsezh0767781.d.html
- [4] Jiemian News, 2026-08-26 — https://www.toutiao.com/article/7678320095768134190/
- [5] NetEase Intelligence, 2026-08-26 — https://www.163.com/dy/article/L59U048505568W0A.html
Reported result: at 1M-token context, the QSA attention kernel achieves up to 7.6x speedup in prefill and 4.9x in decode.
Residual: Gated Residual + FP8
A Gated Residual (GR) mechanism expands the residual stream into 4 parallel branches with dynamic gates controlling read/write. Residual state supports FP8 storage, sharply reducing memory-access overhead—critical for long-context tasks where residual state grows linearly with sequence length.
Embedding: 51B N-gram Embedding
Beyond the 125B main model, a 51B N-gram embedding can be offloaded to host memory, overlapping asynchronous prefetch with computation to avoid occupying GPU VRAM long-term.
Optimization: Muon + Refitted Scaling Law
Training uses the Muon optimizer (from Keller Jordan's research), with optimization of orthogonalization precision, parameter roles, and fused matrix splitting. Scaling laws were re-fitted for the new architecture, since laws fitted on old architectures no longer apply.
Benchmarks: Efficiency Without Capability Trade-offs
| Benchmark | Result | |---|---| | SWE-bench Pro (agentic coding) | Leads Opus 4.6 by 9.1 points | | CoWorkBench (long-horizon professional tasks) | Beats DeepSeek-V4-Flash | | Toolathlon Verified (real tool calls) | Beats DeepSeek-V4-Flash | | JobBench (professional work tasks) | Leads Opus 4.6 by ~20 points | | AndroidWorld (mobile agents) | Beats Opus 4.6 by 22.5 points | | MathVision (visual math reasoning) | Beats Opus 4.6 by 25.1 points | | ERQA (embodied intelligence) | Leads by 31.5 points |
Notably, the pretrained-only Base model reportedly exceeds the 3x-larger Qwen3.7-Plus Base on SuperGPQA (general), GSM8K (math), and SWE-bench Pretrain (coding)—a fundamental capability jump attributed to architecture.
Pricing: Entering the "1-Yuan Era"
| Item | Price | |---|---| | Input | 1 CNY / million tokens | | Output | 3 CNY / million tokens | | vs Claude Opus 4.6 | ~3% | | vs DeepSeek-V4-Flash | As low as 1/3 |
Chinese LLM input price trajectory:
| Time | Model | Input Price (CNY/M tokens) | |---|---|---| | 2024 Q3 | Qwen 2 Max | 40 | | 2025 Q1 | Qwen 2.5 Max | 20 | | 2025 Q4 | DeepSeek V3 | 2 | | 2026 Q1 | Qwen 3 Flash | 1.5 | | 2026 Q3 | Qwen 3.8-Flash | 1 |
That is a 97.5% drop in 18 months, reflecting system-wide cost reduction across architecture, training, inference, and hardware utilization. At 2,000 tokens per conversation, 1 CNY buys roughly 500 chats—effectively removing the cost barrier for consumer AI applications.
Qianwen Office Debuts
Qwen3.8-Flash powers the newly launched Qianwen Office, whose standard mode reportedly covers 95% of daily office tasks: document summarization and translation, data analysis and report generation, PPT outlines, email drafting, meeting minutes and action items, and Excel formulas. These scenarios share a profile of being high-frequency, repetitive, and low-risk. Users need not choose a model—remaining hard cases can switch to Qwen3.8-Max or Pro mode.
Strategically, this marks a shift from "selling API calls" to "shipping native applications"—when marginal model cost is low enough, packaged applications may be more commercially valuable than selling tokens.
Qwen3.8-Flash-Next: The Qwen4 Prototype
Released simultaneously with open weights on Hugging Face and ModelScope, Qwen3.8-Flash-Next previews the next-generation architecture. The workflow: open-source Next → community validation and feedback → build the full Qwen4 family (expected 2026 H2–2027 H1).
As of August 2026, the Qwen3.8 series includes three open-sourced sizes (Qwen3.8-Max, Qwen3.8-27B, Qwen3.8-Flash) totaling 2.4T parameters; 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/M input tokens | | Zhipu | GLM-5.3 (Max / Flash) | Flash open-sourced (320B-A18B), priced at 1/40 of Opus | | DeepSeek | V4 (Pro / Flash / Lite) | V4-Pro live, AgentX load optimization |
Compared internationally:
| Model | Total / Active | Input Price (CNY/M tokens) | |---|---|---| | Claude Opus 4.8 | Unknown | ~30 | | GPT-5 | Unknown | ~25 | | Qwen3.8-Flash | 125B / 6B | 1 | | DeepSeek-V4-Flash | 238B / 21B | 3 | | GLM-5.3-Flash | 320B / 18B | 1.5 (limited time) |
The 30x price gap signals not a movement down the same price curve but a new inclusive price tier—enabling AI adoption in developing markets, SMEs, consumer apps, and long-tail use cases, while pressuring closed-source, high-price business models.
A Underestimated Risk: Can Price Wars Sustain R&D?
Each price cut is backed by engineering upgrades that require heavy R&D investment. If revenue growth lags model iteration, R&D could be squeezed. The post argues Qwen3.8-Flash offers a sustainable path: architecture upgrades (sparse activation, attention optimization, Muon, refitted scaling laws) reduce per-inference cost while total parameters—and capability—keep growing. This is framed as an asymmetric competitive strategy against closed-source incumbents.
Outlook: 12 Months Ahead
1. Full Qwen4 family: based on the Next architecture, expected 2026 H2–2027 H1, possibly with Max / Pro / Flash / Lite tiers. 2. Native multimodality expansion: toward video understanding, video generation, and 3D understanding. 3. Agent deepening: agentic capability (e.g., the SWE-bench Pro lead) as a Qwen differentiator.
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 | | Pricing | 1 CNY in / 3 CNY out per million tokens | | Context | Native 262K, up to 1M via YaRN | | Architecture | QSA + GDN + Gated Residual + 51B N-gram Embedding + Muon | | Qwen4 prototype | Qwen3.8-Flash-Next (open weights) | | Qwen3.8 series total params | 2.4T | | Qwen downloads | 3B+ / 300K+ derivatives |