Qwen releases open weights for Qwen3.8-2.4T-A95B, a 2.4T-parameter MoE with 95B activated

The repository publishes weights and configuration files for Qwen3.8-2.4T-A95B in Hugging Face Transformers format, which the project describes as the first Qwen-Max-class model made available as an open release. Only one feed carried the event and the Hacker News entry points to a comments thread rather than independent reporting, so the available signal is essentially a single primary source. Claims about gains in coding, professional work, research, and long-horizon agentic tasks therefore come from the publisher's own benchmark table and should be read accordingly.

Qwen3.8-2.4T-A95B is a causal language model with 2.4T total parameters and 95B activated per token, structured as 92 layers using the layout 23 × (3 × (Gated DeltaNet → MoE) → 1 × (Gated Attention → MoE)). The MoE has 512 experts with 10 routed plus 1 shared active per token, an expert intermediate dimension of 2048, and a hidden dimension of 8192 with a 248,320-vocabulary padded embedding. The Gated DeltaNet block uses 128 V heads and 16 QK heads at head dimension 128, while Gated Attention uses 64 Q heads and only 4 KV heads at head dimension 256, an asymmetric KV layout that materially affects KV-cache memory and is worth budgeting for at long context.

Native context is 262,144 tokens and the card states the model is extensible to 1,010,000 tokens. Training included multi-token prediction across multiple steps, and reasoning is surfaced as two runtime controls: reasoning_effort tunes how much the model thinks, while preserve_thinking retains reasoning context across historical messages. Both controls are directly relevant to building agents that need consistent chain-of-thought across many turns rather than a single short prompt.

The artifact is positioned as a base rather than a finished product. The weights target Hugging Face Transformers and are stated to be compatible with vLLM, SGLang, and TokenSpeed, so existing inference stacks can load the model without a new runtime. The official Qwen3.8-Max, built on Qwen3.8-2.4T-A95B, is the variant that adds vision input, non-thinking support, 1M default context, and built-in tools, and is delivered through the managed Qwen API. That split means open-weights users get the foundation while the tool-augmented experience stays on the hosted service.

A 2.4T-total-parameter MoE is beyond what a single workstation can hold, even with only 95B parameters active per token, and adopting it locally requires multi-node GPU-class hardware. Coverage here is limited to a single feed plus the model card, so benchmark numbers are publisher-reported and the table's caveat that some Fable 5 results may involve fallbacks should be carried into any comparison. Finally, because Qwen3.8-Max is the feature-complete variant, the open artifact is explicitly a starting point for downstream systems rather than a drop-in replacement for the hosted product.