Alibaba Releases Qwen3.8-2.4T-A95B-FP8: 2.4T-Parameter Open Model with 1M-Token Context

Alibaba Releases Qwen3.8-2.4T-A95B-FP8

Alibaba's Qwen team has published Qwen3.8-2.4T-A95B-FP8, an FP8-quantized, open-weight checkpoint of its largest model to date. The release is available now on Hugging Face in Transformers format, with compatibility for vLLM, SGLang, and TokenSpeed inference stacks.

The model is a Mixture-of-Experts (MoE) architecture with 2.4 trillion total parameters and 95 billion activated per forward pass. It uses 512 total experts, activating 10 routed experts plus 1 shared expert per token, an intermediate dimension of 2048, and 92 layers built on a hybrid Gated DeltaNet / Gated Attention block structure. The model was trained with multi-token prediction (MTP).

Context length is 262,144 tokens natively, extensible up to 1,010,000 tokens. Alibaba says the FP8 quantization — using fine-grained blocks of 128 — produces performance "nearly identical" to the full-precision original, according to the company.

This checkpoint serves as the open-weight foundation for Qwen3.8-Max, the fully-featured hosted version available through Qwen Cloud's API, which adds vision input, non-thinking mode, a default 1M-token context, and built-in tools. Pricing for Qwen Cloud API access was not disclosed in the release documentation.

Benchmark Claims

According to Alibaba's published benchmarks, Qwen3.8-Max scores competitively against Claude Opus 4.8, Claude Fable 5, and GPT-5.6 Sol across coding and agentic tasks:

• Terminal Bench 2.1: 86.6 (vs. Opus 4.8's 84.6, GPT-5.6 Sol's 88.8)
• SWE-bench Pro: 67.7 (vs. Opus 4.8's 69.2, Fable 5's 80.0)
• PaperBench: 93.0 (vs. Fable 5's 88.8, GPT-5.6 Sol's 90.5)
• GPQA Diamond: 92.6 (matching Fable 5, behind GPT-5.6 Sol's 94.1)
• HLE (Humanity's Last Exam): 43.6 (below Fable 5's 53.3 and GPT-5.6 Sol's 47.2)

Compared to its predecessor Qwen3.7-Max, the new model shows substantial gains — for example, DeepSWE 1.1 jumps from 21.6 to 56.6, and JobBench rises from 31.3 to 53.4. These figures come from Alibaba's own evaluation harnesses, largely using Claude Code as the test framework, and have not been independently verified.

The model introduces "flexible thinking control" via a reasoning_effort parameter and a preserve_thinking flag that retains reasoning context across multi-turn conversations, positioning it explicitly as a reasoning-capable model for long-horizon agentic work.

What this means

This is Alibaba's first release of a Qwen-Max-class model with open weights, narrowing the gap between what the company sells commercially and what it distributes for self-hosting. At 2.4T total parameters, running this model requires serious infrastructure even with only 95B activated per token and FP8 quantization — this is not a model for consumer hardware. The benchmark comparisons against Opus 4.8, Fable 5, and GPT-5.6 Sol suggest Alibaba is positioning Qwen3.8 as a frontier-tier coding and agentic model, though all reported scores come from Alibaba's own testing setup and should be treated as claims pending independent replication. The extended 1M-token context and MTP training also signal continued focus on long-horizon agent workloads rather than short single-turn chat performance.