Alibaba's Qwen team releases Qwen3.8-2.4T-A95B, a 2.4-trillion-parameter open-weight model

Alibaba's Qwen team has published weights for Qwen3.8-2.4T-A95B on Hugging Face, a mixture-of-experts language model with 2.4 trillion total parameters and 95 billion active per token. The model card calls it "the most capable generation in the Qwen open-model family to date," positioning it as Qwen's new flagship ahead of a crowded field of large Chinese open-weight releases from DeepSeek and Z.ai.

What's new

• Scale: 2.4T total parameters, 95B active per token, routed through 512 experts with 11 activated per token (10 routed plus 1 shared).
• Architecture: A hybrid design of 23 repeated blocks, each combining 3 Gated DeltaNet layers with 1 Gated Attention layer, integrated with the MoE routing layers; 92 layers total and an 8,192 hidden dimension.
• Context: 262,144 tokens natively, extendable to 1,010,000 tokens.
• Reasoning: Text-only, requires "thinking mode" for all interactions, with an adjustable reasoning_effort parameter and Multi-Token Prediction used in training.
• Benchmarks: GPQA Diamond 92.6, SWE-bench Pro 67.7, Deep SWE 1.1 56.6, Terminal Bench 2.1 86.6 — all weighted toward coding and long-horizon agentic tasks.
• License: A custom "qwen3.8-max" license rather than a standard permissive license like Apache 2.0, though the safetensors weights are directly downloadable via Hugging Face Transformers.

Context

Qwen has kept up an aggressive release cadence through 2026, iterating from the 3.6 series (including a 27B dense model) into this larger MoE flagship. The release lands alongside a wave of frontier-scale open-weight competition: DeepSeek has been pushing V4-Pro and V4-Flash toward general availability, and Z.ai's GLM-5.2 already offers a 1M-token context window. Qwen3.8-2.4T-A95B's spec sheet — massive total parameter count, modest active-parameter footprint, million-token extended context — reads as a direct answer to both. The model card recommends managed inference through Qwen Cloud for production workloads, suggesting the raw weights are aimed more at researchers and self-hosters than at casual deployment given the compute required to serve a 2.4T-parameter MoE model.

Why it matters

The release keeps up the pattern of open-weight frontier-scale models emerging from Chinese labs at a pace that increasingly tracks closed labs' own releases, with agentic coding benchmarks like SWE-bench Pro now landing in a range that's competitive with proprietary frontier models. The custom, non-OSI license is also notable: as with several other recent "open-weight" flagship releases, Qwen is publishing usable weights without adopting a fully permissive open-source license, a distinction that matters for enterprises evaluating redistribution and fine-tuning rights before adopting the model.