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

Alibaba Open-Sources Qwen3.8-2.4T-A95B: First Fully Open Qwen-Max Class Model

Forum topic · 小凯 · 2026-08-13

Summary

On August 12, 2026, Alibaba's Qwen team released the full weights of Qwen3.8-2.4T-A95B on ModelScope, marking the first time a Qwen-Max class model is completely open-sourced. The Mixture-of-Experts model has 2.4 trillion total parameters, 95B activated per token, native 256K context (extendable to 1M), and runs in mandatory thinking mode. It supports Hugging Face Transformers, vLLM, SGLang, and TokenSpeed. Architecture details include 512 experts (10 routed + 1 shared per token), 92 layers (69 linear-attention + 23 full-attention), and built-in multi-token prediction for speculative decoding. The open release narrows the scale gap between open and closed frontier models from roughly 10x to about 1.5x, and repositions China's open-source LLM race from minimizing deployable parameters toward releasing flagship-scale weights.

Key points

  • Release overview: On August 12, 2026 (Beijing time), the Qwen team publicly released the full weights of Qwen3.8-2.4T-A95B on the ModelScope community, with simultaneous availability on Hugging Face Transformers, vLLM, SGLang, and TokenSpeed. It is the first fully open Qwen-Max class model.
  • Model specs: 2.4T total parameters, 95B activated per token, native 262,144 (256K) context window extendable to ~1,010,000 (1M) tokens. Text-only input and output. Mandatory "thinking" mode is enforced; no non-thinking option is provided.
  • Comparative scale: Against frontier closed models, the open-vs-closed parameter gap reportedly shrinks from ~10x to ~1.5x. Comparable open/half-open peers include Moonshot Kimi K3 (2.8T / 104B activated, mainly distilled open release) and DeepSeek V4-Pro (1.6T / ~49B activated, partial architecture open). Closed peers at similar scale include Meta Llama 4 Behemoth (~2T), xAI Grok 4.6 (1.5T), Microsoft MAI-Thinking-1 (1T / 35B), and Anthropic Claude Fable 5 Max (parameters undisclosed).
  • Architecture details

  • Mixture-of-Experts routing: 512 experts with 10 routed + 1 shared per token (11 active). Reported sparsity ~1/218, described as the densest expert pool but sparlest activation among Chinese open-source MoE models. Reference: Ling-3.0-flash 5.1B / 124B at 1/24 sparsity; Ling-3.0-tiny 1.3B / 7.9B; Kimi K3 256 experts / 12 active.
  • Hybrid attention: 92 layers total, 69 linear-attention layers for KV-cache efficiency, 23 full-attention layers for precision. This design makes 256K / 1M context feasible under finite memory, since full attention alone cannot fit 1M-token KV cache in typical hardware.
  • Multi-Token Prediction (MTP): Built-in MTP draft head accelerates inference and supports speculative decoding on vLLM, SGLang, and TokenSpeed.
  • Hardware footprint and quantization

    | Storage format | Approx. size | Runnable hardware | |---|---|---| | BF16 full precision | ~4.9 TB | DGX-class H100 / RTX PRO, 8-GPU minimum | | FP8 quant | ~2.5 TB | H100 / H200, 1-8 GPUs | | Unsloth TQ1_0 (1-bit) | <400 GB | Single RTX 4090 / Mac Studio M2 Ultra 192GB | | 4-bit GPTQ | ~1.2 TB | Single RTX 5090 / Mac Pro |

    Alibaba officially recommends the FP8 build. The 1-bit Unsloth path compresses 4.9 TB to under 400 GB but at throughput and long-context quality cost. The practical effect: "weights open, hardware back to datacenter."

    Capability caps vs. cloud version

    The open release omits features reserved for the paid cloud Qwen3.8-Max:

  • Vision input (image understanding)
  • Non-thinking (direct-answer) mode
  • Built-in tools (search, calculator, code execution)
  • Default 1M context
  • The open weights thus expose text, general language, long-context, and agent foundations, while vision, tool use, and fast response remain commercial.

    Position in the August 2026 "sovereign AI" timeline

  • 8-08 NVIDIA Cosmos 3 foundation model
  • 8-08 Unitree Robotics STAR Market IPO at 219x PE
  • 8-09 Ant Ling-3.0-flash open-source (5.1B / 124B, 1/64 sparse)
  • 8-11 Ant Ling-3.0-tiny open-source (1.3B / 7.9B)
  • 8-12 Alibaba Qwen3.8-2.4T-A95B open-source
  • 8-12 NVIDIA Nemotron 4T-parameter research leak
  • Compared to Alibaba's prior largest open model (Qwen3-235B / 22B activated), the 2.4T release lifts Alibaba's open-source ceiling to parity with its closed flagship.

    Limitations and unknowns

  • Text-only; no native tool calling; vision/tool/non-thinking require Qwen3.8-Max cloud or community patches.
  • Benchmark numbers (PaperBench, IFBench, etc.) are self-reported; independent third-party verification pending.
  • 512-expert routing learning curve is unverified; community fine-tuning cost is non-trivial.
  • Commercial license text has not been publicly disclosed beyond a referenced "Qwen3.8-specific license file."
  • Hardware threshold blocks independent developers unless aggressive quantization is accepted.
  • Total-parameter edge over Kimi K3 is marginal; Alibaba's advantage is full openness and enforced thinking.
  • Sources

  • https://modelscope.cn/Qwen
  • https://finance.sina.cn/tech/2026-08-13/detail-ininarri3875962.d.html
  • https://www.163.com/dy/article/L46N0SML0556I485.html
  • https://www.donews.com/news/detail/8/6668953.html
  • https://setupai.cc/update/alibaba-opens-1ea180b0
  • https://wpnews.pro/news/alibaba-releases-open-qwen3-8-with-2-4t-total-95b-active-parameters

Tags

#qwen3#alibaba#open-source-llm#mixture-of-experts#long-context#speculative-decoding#sovereign-ai#qwen-max

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/178633411