On August 12, Alibaba's Qwen team officially open-sourced the weights of Qwen3.8-2.4T-A95B — 2.4 trillion total parameters, 95B activated per token, 512 experts with 11 activated per inference, and 92 layers of hybrid attention. Native context is 262,144 tokens, extendable to roughly 1 million via YaRN, with a recommended max output of 131,072 tokens. This is the first time Alibaba has fully released weights of a Qwen-Max-class flagship — not a distillation, not a pruned MoE, not a truncated version, but the open-source equivalent of Qwen3.8-Max.
MoE Architecture: How 2.4T "Looks Like" Only 95B
The "A95B" in the name is the answer — Activation 95B, meaning each token activates about 95 billion parameters. Architecture details:
- Total parameters: 2.4 trillion
- Activated parameters: ~95B per token
- Number of experts: 512
- Experts activated per inference: 11
- Layers: 92 layers of hybrid attention (MLA + sliding-window attention mix)
- Native context: 262,144 tokens
- Extended context: ~1,010,000 tokens via YaRN
- Max output: 131,072 tokens
- Terminal Bench 2.1: Qwen3.8 86.6 / GLM-5.3 88.2 / DeepSeek V4 Pro 87.9 / Fable 5 88.0
- DeepSWE v1.1: Qwen3.8 56.6 / GLM-5.3 66.9 / DeepSeek V4 Pro 62.7 / Fable 5 67.5
- CyberGym (cybersecurity): Qwen3.8 78.1 / GLM-5.3 84.5 / Fable 5 83.8
SiliconFlow launched it Day-0, priced at $2.00/million input tokens and $6.00/million output tokens — nearly identical to Grok 4.6's $2/$6.
However, the open-weight version differs from the commercial API version: the open release is text-only (no vision or other multimodal input), and thinking mode is forced on (every response first generates a chain of thought, with no way to disable it). For vision input and non-thinking mode, you need the Qwen3.8-Max commercial API.
Day-0 on Domestic Compute: 9 AI Chips Adapted Simultaneously
Per the GitHub AI Daily report, Qwen3.8-2.4T-A95B achieved Day-0 adaptation on 9 domestic Chinese AI chips on release day. Mainstream domestic chips — Cambricon, Hygon, Ascend, Enflame, Iluvatar (Days), Moore Threads, Biren, etc. — can run this 2.4T model immediately, without waiting for community ports.
This matters more than the parameter scale itself. Previously, domestic open-source models typically took 2–4 weeks to run on Ascend/Hygon/Cambricon. Day-0 adaptation means:
1. Compute scheduling returns to domestic hands. DeepSeek V4 Pro and GLM-5.3 previously ran on NVIDIA first with domestic support added later; Qwen3.8 is the first flagship usable on domestic compute from day one. 2. Shorter enterprise deployment paths. Government and state-owned enterprise customers can go from "waiting for domestic adaptation" to "procure domestic chips and deploy at launch." 3. Price war arrives early. SiliconFlow's Day-0 launch plus 9-chip adaptation means Qwen3.8 inference costs should quickly fall to $1/$3 or lower.
Coding: Strong Reasoning, Weak on Long-Horizon Coding
Qwen3.8 performs strongly on reasoning benchmarks: GPQA Diamond 92.6%, PaperBench 93.0%, Terminal-Bench 2.1 86.6%. But on DeepSWE v1.1 (long-horizon software engineering) it scores only 56.6, clearly behind GLM-5.3's 66.9 and DeepSeek V4 Pro's 62.7.
Coding agent comparison:
How to Choose: Qwen3.8 vs GLM-5.3 vs DeepSeek V4 Pro
| Model | Input/Output (USD/million) | Notes | |---|---|---| | Qwen3.8-2.4T-A95B (SiliconFlow) | $2.00 / $6.00 | Open-sourced; Day-0 on 9 chips | | GLM-5.3 (API) | TBD | Open weights in two weeks | | DeepSeek V4 Pro (pre-8/17 peak) | $1.27 / $3.80 | Still the cheapest flagship globally | | Grok 4.6 (>200K billed at higher rate) | $2.00 / $6.00 | Over 200K tokens, entire request billed at $4/$12 |
One-line guidance: for open-source self-hosting, choose Qwen3.8 (available now) or wait for GLM-5.3 (two weeks); for the strongest open-source coding, choose GLM-5.3; for cheap and capable, choose DeepSeek V4 Pro (lock in pricing before the 8/17 increase); for long-running agents + vision, choose Grok 4.6.
Three signals to watch over the next month: 1) whether chip vendors publish Qwen3.8 inference benchmarks (especially Ascend 910C); 2) the number of LoRA/QLoRA fine-tuned derivatives produced by the community within 24 hours; 3) whether the capability gap between Qwen3.8-Max API and the open version continues to shrink (distillation loss curve).
Sources: Qwen GitHub, ModelsLab Qwen3.8-2.4T docs, SiliconFlow announcement on X, ModelScope community Qwen3.8 page, GitHub AI Daily 8-15.