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

Alibaba releases Qwen3.8-2.4T-A95B: 2.4T MoE flagship goes open-weight with Day-0 support on 9 Chinese AI chips

Forum topic · 小凯 · 2026-08-15

Summary

On August 12, Alibaba open-sourced Qwen3.8-2.4T-A95B, a 2.4-trillion-parameter Mixture-of-Experts model that activates 95B parameters per token using 11 of 512 experts across 92 mixed-attention layers. It is the first full open-weight release of a Qwen-Max-class flagship, supporting a native 262,144-token context, extendable to about 1 million tokens via YaRN, with recommended max output of 131,072 tokens. SiliconFlow launched Day-0 hosting at $2.00/$6.00 per million input/output tokens, matching Grok 4.6 pricing. The same day, nine Chinese AI accelerators from Cambricon, Hygon, Ascend, Enflame, Tianshu, Moore Threads, Biren and others shipped Day-0 support, shortening enterprise deployment paths. Benchmarks show strong reasoning (GPQA Diamond 92.6%, Terminal-Bench 2.1 86.6%) but weaker long-horizon coding (DeepSWE v1.1 56.6 vs GLM-5.3 66.9). Open weights are text-only with forced thinking mode.

Alibaba has open-sourced Qwen3.8-2.4T-A95B, a 2.4-trillion-parameter Mixture-of-Experts (MoE) large language model and the first full open-weight release of a Qwen-Max-class flagship. SiliconFlow launched Day-0 hosting, and nine Chinese AI chips shipped Day-0 compatibility.

MoE Architecture: 2.4T That "Looks Like" 95B

The "A95B" suffix denotes Activation 95B: roughly 95 billion parameters are activated per token.

  • Total parameters: 2.4T
  • Activated parameters: ~95B/token
  • Experts: 512
  • Active experts per token: 11
  • Layers: 92 hybrid-attention layers (MLA + sliding-window attention)
  • Native context: 262,144 tokens
  • Extended context: ~1,010,000 tokens via YaRN
  • Max output: 131,072 tokens
  • SiliconFlow Day-0 Hosting

    SiliconFlow launched the model on day one at $2.00 per million input tokens and $6.00 per million output tokens, roughly matching Grok 4.6's $2/$6. The open-weight release differs from the commercial Qwen3.8-Max API: the open version is text-only, has no multimodal (vision) input, and forces thinking mode on for every reply, with no toggle to disable. Vision input and non-thinking mode require the proprietary Qwen3.8-Max API.

    Domestic Compute: Day-0 on 9 Chinese AI Chips

    GitHub AI Daily confirmed that on launch day, nine Chinese AI chips completed Day-0 adaptation, including Cambricon, Hygon, Ascend, Enflame, Tianshu, Moore Threads, Biren, and others. Previously, domestic open-source models typically took 2-4 weeks to run on Ascend/Hygon/Cambricon silicon. Implications:

    1. Compute scheduling returns to domestic vendors. Qwen3.8 is the first flagship to be "domestic-compute-ready at launch" rather than launching on NVIDIA first. 2. Shorter enterprise deployment cycles. Government and SOE customers can procure domestic chips and deploy immediately on release. 3. Earlier pricing pressure. SiliconFlow Day-0 plus 9-chip adaptation is expected to push inference costs toward $1/$3 or lower.

    Coding: Strong Reasoning, Weaker Long-Horizon Engineering

    Qwen3.8 is competitive on reasoning benchmarks: GPQA Diamond 92.6%, PaperBench 93.0%, Terminal-Bench 2.1 86.6%. However, on DeepSWE v1.1 (long-horizon software engineering) it scores only 56.6, trailing GLM-5.3 (66.9) and DeepSeek V4 Pro (62.7).

    Coding-agent comparison:

  • 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: Qwen3.8 78.1 / GLM-5.3 84.5 / Fable 5 83.8
Qwen3.8's strength is breadth of reasoning (math, paper reproduction, OS-level operations); its weakness is end-to-end multi-file project authoring. This implies an effective 10-15% gap between open-weight and commercial Qwen3.8-Max.

How to Choose: Qwen3.8 vs GLM-5.3 vs DeepSeek V4 Pro

| Model | Input/Output (USD per 1M tokens) | Notes | |---|---|---| | Qwen3.8-2.4T-A95B (SiliconFlow) | $2.00 / $6.00 | Open weights, Day-0 on 9 Chinese chips | | GLM-5.3 (API) | TBD | Open weights in ~2 weeks | | DeepSeek V4 Pro (before 8/17 peak) | $1.27 / $3.80 | Still the cheapest flagship globally | | Grok 4.6 (>200K rebills) | $2.00 / $6.00 | $4/$12 for segments over 200K tokens |

Quick guidance: choose Qwen3.8 for open-weight self-hosting today, GLM-5.3 in ~2 weeks for the strongest open-source coding, DeepSeek V4 Pro for the lowest price (lock rates before the August 17 hike), and Grok 4.6 for long-horizon agents plus vision.

Signals to Watch in the Next Month

1. Whether domestic chip vendors publish Qwen3.8 inference benchmarks, especially Ascend 910C. 2. The count of LoRA/QLoRA fine-tunes the community produces within 24 hours. 3. Whether the gap between Qwen3.8-Max commercial API and the open-weight version narrows (distillation-loss curves).

Sources: Qwen GitHub, ModelsLab Qwen3.8-2.4T docs, SiliconFlow X announcement, ModelScope Qwen3.8 page, GitHub AI Daily 8-15.

Tags

#qwen3.8#alibaba#moe#open-source-llm#chinese-ai-chips#siliconflow#llm-benchmarks#ai-inference

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/178633490