News
ResearchOpenAI pens AI safety agreement with four firmsReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierOpenAI鈥檚 New Hacking Model Found Dozens Of Holes In Apple TechFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIResearchOpenAI pens AI safety agreement with four firmsReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierOpenAI鈥檚 New Hacking Model Found Dozens Of Holes In Apple TechFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAI
EveryLLM
Qwen3.8-2.4T-A95B

Qwen3.8-2.4T-A95B

Open-weight reasoning model with 2.4 trillion parameters

Qwen3.8-2.4T-A95B is a large language model released by the Qwen team in 2026. It features 2.4 trillion total parameters with 95 billion active per token. The model focuses on complex reasoning, coding, and agentic tasks using a hybrid architecture.

Share on WhatsApp

  • 2.4 trillion total parameters with 95 billion active
  • Hybrid architecture using Gated DeltaNet and MoE
  • Native context window of 262,144 tokens
  • Text-only model requiring enabled thinking mode
  • Optimized for long-horizon agentic and coding tasks

What it is

Qwen3.8-2.4T-A95B is a post-trained language model from the Qwen family. It represents the most capable generation in the open-model series as of 2026. The model contains 2.4 trillion parameters in total, with 95 billion activated during inference. It is designed for high-performance tasks in coding, professional work, and research. The system excels at long-horizon agentic tasks that require multiple steps to complete. Unlike some predecessors, this specific variant is text-only and requires thinking mode for all interactions.

How it came about

The model follows the Qwen3.5 and Qwen3.6 series, which saw widespread community adoption. Qwen3.8 was introduced as a significant upgrade, bringing Qwen-Max-class capabilities to an open release for the first time. It builds directly on the architectural foundation laid by Qwen3.5. The release aims to provide substantial gains in reliability for complex, multi-step tasks. This version marks a shift towards offering top-tier performance in an accessible, open-weight format.

Inside the model

The architecture uses a hybrid design combining Gated DeltaNet and Gated Attention mechanisms within a Mixture of Experts (MoE) framework. Each block contains three layers of Gated DeltaNet followed by one layer of Gated Attention. The model activates 10 routed experts and one shared expert per token. It supports a native context length of 262,144 tokens, extendable to over one million. Multi-token prediction was used during training. The tokenizer has a padded embedding size of 248,320. Reasoning depth can be adjusted via parameters, and historical reasoning context is preserved by default.

What it can do

The model shows strong performance in autonomous planning and handling environment feedback. It is optimized for end-to-end task completion in coding and software engineering. Benchmarks include evaluations on Terminal Bench, SWE-bench Pro, and various in-house tests like QwenSWEBench. It performs well in professional domains such as finance, law, and medicine through the CoWorkBench evaluation. The system allows tuning of reasoning effort from low to extra high, balancing speed and accuracy based on task complexity.

How to use it

Weights are available openly in Hugging Face Transformers format. They are compatible with inference engines like vLLM, SGLang, and TokenSpeed. For managed services, the model is offered via the Qwen Cloud API as Qwen3.8-Max. This hosted version includes additional features like vision input and built-in tools. The open-weight version requires users to manage their own infrastructure but allows for flexible deployment and integration into existing development stacks.

Questions people ask about Qwen3.8-2.4T-A95B

Is this model multimodal?

No, Qwen3.8-2.4T-A95B is a text-only model. It does not support image or video inputs. For multimodal capabilities, users should look to other variants in the Qwen3.8 family, such as the 27B version, which includes native vision understanding.

Can I disable the reasoning process?

No, thinking mode cannot be disabled for this specific model. Every response automatically begins with reasoning enclosed in special tags before the final output. This is a core design feature intended to improve reliability on complex tasks.

What hardware is needed to run it?

With 95 billion active parameters, it requires significant GPU memory. However, the mixture of experts design means only a fraction of the total 2.4 trillion parameters are active at once. Efficient inference frameworks like vLLM or SGLang are recommended for deployment.

How does it handle long documents?

The model supports a native context length of 262,144 tokens. This can be extended up to 1,010,000 tokens. This makes it suitable for processing large codebases, long reports, or extensive historical conversation logs without losing context.

Good reading and sources

    Where this page comes from

    There is no English Wikipedia article on Qwen3.8-2.4T-A95B yet, so the facts on this page come from the sources listed above: huggingface.co/Qwen/Qwen3.8-2.4T-A95B, huggingface.co/Qwen/Qwen3.8-27B. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.

    It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.

    The weekly digest

    Every Monday: the model of the day, new pages on EveryLLM, and the conferences and anniversaries of the week ahead.