Qwen3.8 marks the first open release of a Qwen "Max"-class model optimized to carry complex, multi-step tasks through to completion rather than only answer isolated prompts. The practical difference is its design focus on agent execution and sustained reasoning: it preserves internal reasoning context, exposes adjustable reasoning depth, and allocates large internal token budgets so the model can plan, react to environment feedback, and produce long-form final outputs within a single session.
Key Capabilities
- Architecture and scale: mixture-of-experts causal LM with 2.4T total parameters and 95B activated capacity; 92 layers with a hybrid Gated DeltaNet + Gated Attention layout and 512 experts (10 routed + 1 shared active experts). This layout targets high effective capacity while keeping per-token compute constrained via MoE routing.
- Long-context and multi-token planning: native 262,144-token context (extensible up to 1,010,000 tokens) and Multi-Token Prediction (MTP) training to enable long-horizon planning, internal chain-of-thought style reasoning, and large final outputs.
- Reasoning controls: exposes reasoning_effort levels (xhigh/medium/low) and preserve_thinking behavior to tune internal analysis depth and cost; Qwen3.8-2.4T-A95B is text-only and enforces thinking-mode (responses begin with reasoning wrapped in
<think>...</think>by default). - Agent and coding focus: tuned for autonomous planning, tool/feedback handling, and coding-agent workflows; benchmarked across many internal and external suites (coding, agentic benchmarks, domain-specific tests) to position it for sustained multi-step tasks.
- Deployment compatibility: provided artifacts in Transformers/safetensors format and tested with serving engines such as vLLM, SGLang, and TokenSpeed; recommended sampling defaults and separate token budgets for internal reasoning vs final responses.
Who it's for and tradeoffs
Great fit if you need a model that must maintain long internal reasoning traces and produce very long, coherent final outputs (e.g., long-running autonomous agents, complex coding agents, multi-stage research assistants). It also suits teams prepared to deploy MoE models with specialized inference engines and large memory/throughput resources.
Look elsewhere if you require a lightweight, multimodal, or thinking-free model for low-latency consumer chat; Qwen3.8-2.4T-A95B is text-only and requires thinking-mode for all interactions. Operational trade-offs include higher infrastructure complexity (MoE routing, memory and compute demands), potential licensing constraints (model card lists license: other), and the need to use supported high-performance serving frameworks for cost-effective inference.
Practical notes
- Not multimodal: this Hugging Face artifact is text-only; Qwen3.8-Max (commercial offering) adds vision & additional features.
- Inference tips: recommended sampling settings and guidance for allocating large token budgets for internal reasoning vs final response are provided in the model card; production deployments typically use vLLM/SGLang/TokenSpeed or Qwen Cloud for managed inference.
Overall, Qwen3.8-2.4T-A95B is positioned as an open-release, Max-class LLM that prioritizes sustained reasoning and agentic workflows at very large context and capacity scales; adopt it when you can support MoE operational complexity and need long-horizon, multi-step task completion.
