Qwen (Alibaba) has released the open weights for the Qwen3.8-2.4T-A95B model — this is the first Qwen-Max-level model with a fully open architecture. 2.4 trillion parameters, MoE with 512 experts, a context of up to one million tokens, and day-0 support for vLLM and SGLang make the model immediately ready for deployment.
What happened
Alibaba has made the weights of the Qwen3.8-2.4T-A95B model open. The model uses a hybrid architecture of 92 layers of Gated DeltaNet and Gated Attention with a mixture of experts: out of 2.4 trillion parameters, 95 billion are active at a time with 512 experts. The context window is up to 1 million tokens. Weights are available in BF16 (about 5 TB), FP8 (about 2.4 TB), and quantized versions NVFP4 and MXFP4 (about 1.2 TB). In benchmarks, the model showed AIME25 at a level of 92–96% and PaperBench 93.0, taking leading positions in most multimodal and agentic tasks. The vLLM and SGLang frameworks added support from the first day of the weights release.
Context
Prior to this release, frontier models with trillions of parameters were available primarily through closed APIs: GPT-5.6 Sol from OpenAI, Gemini 3.1-Pro from Google, Fable 5, and Opus 4.8. Open models with hundreds of billions of parameters existed, but a 2T+ parameter class model with open access is a precedent. The architectural choice of Gated DeltaNet + Gated Attention is significant: linear attention in most layers gives O(1) memory complexity per layer instead of O(N) for standard self-attention, which is critical for working with a 1M token context. The MoE configuration with a ratio of 95B out of 2.4T (about 37% active) is close to the optimal balance between expert specialization and routing overhead.
Why this matters for the industry
This is a direct challenge to closed frontier models: performance at the level of GPT-5.6 Sol and Opus 4.8 is now available for local deployment and modification without being tied to an API provider. Inference infrastructure is ready — vLLM and SGLang support the model from day one. For teams with access to NVIDIA B300 or AMD MI355X clusters, inference on a single node is possible in FP4. The Gated DeltaNet + MoE architecture could become a standard pattern for efficient 1T+ parameter models, as linear memory for context solves the key bottleneck limitation for long workloads.
Why this matters for users
If you have access to a single B300 or MI355X node, you can download the FP4 version and run inference via vLLM or SGLang. For multi-node clusters, FP8 and BF16 versions are available. Recommended generation parameters: temperature=1.0, top_p=0.95, max_tokens of at least 128_000 for reasoning tasks. Teams without access to flagship hardware can expect quantized ports for consumer GPUs. Product teams can start integrating into agent workflows and A/B testing against current models today.
What is still unknown / limitations
The claim of superiority or parity with GPT-5.6 Sol, Gemini 3.1-Pro, and Opus 4.8 is based on Alibaba's self-reports and has not yet been confirmed by independent verification. The AIME25 result is given as a range of 92–96%, not an exact value. Methodological details of training — including the volume and composition of data, training schedule, and cluster configuration — are not fully disclosed. Independent benchmark replications will take 1 to 2 weeks. Real performance on production data may differ from benchmark results.
Sources
- Qwen3.8-2.4T-A95B — model page on Hugging Face
- Day 0 Support for Qwen3.8-2.4T-A95B on vLLM — vLLM blog
- Qwen3.8 — SGLang documentation and recipes
Author
Look at AI, editorial team