Qwen3.8 goes open: a 2.4T-parameter, 95B-active MoE built for agentic coding

Alibaba’s Qwen team has published Qwen3.8-2.4T-A95B, billing it as the strongest model in its open lineup and the first time a “Qwen-Max-class” system has shipped as downloadable weights rather than staying behind an API. Architecturally it’s a Mixture-of-Experts model with 2.4 trillion total parameters but only about 95 billion active per token, drawing on a pool of 512 experts (10 routed plus one shared). The 92-layer stack is a hybrid design that interleaves linear-attention Gated DeltaNet blocks with conventional Gated Attention, adds multi-token prediction, and handles 262K tokens of context natively, extensible past one million.

The release is pitched squarely at coding and long-horizon agent work: autonomous planning, more reliable multi-step task completion, and broad compatibility with existing developer harnesses. Reasoning depth is tunable through a reasoning_effort setting, with prior thinking retained via preserve_thinking. The weights arrive in Hugging Face Transformers format and run on vLLM, SGLang, and Docker, while a hosted Qwen3.8-Max variant on Qwen Cloud layers on vision input, a default 1M-token context, and built-in tools.

The strategic tell is in the benchmark table. Nearly every result — SWE-bench Pro, Terminal Bench, FrontierSWE, and a raft of in-house suites — is run through Anthropic’s Claude Code harness and measured against frontier systems like Claude Opus 4.8, Fable 5, and GPT-5.6. These are vendor-reported figures that warrant independent replication, but the signal is unambiguous: a frontier-scale, openly available MoE tuned for the agentic-coding territory that closed API models have largely owned.