Qwen3.8 2.4T: How Alibaba's Open-Weights MoE Model Matches Frontier AI at 4% Activation

Alibaba's Qwen team has just dropped Qwen3.8-2.4T-A95B, and it changes the math on what open-weights AI can do. At 2.4 trillion total parameters with only 95 billion activated per token, this Mixture-of-Experts model matches — and in several benchmarks exceeds — frontier proprietary models like Claude Opus 4.8, Fable 5, and GPT-5.6 Sol. And anyone can download it.

What Makes Qwen3.8 Different

The Qwen3.8 release is significant for several reasons, but the headline is simple: this is the first time a Qwen-Max-class model has been released with open weights. Previous Qwen-Max models were available only through Qwen Cloud, Alibaba's managed API service. Now the post-trained weights are on Hugging Face, ready for anyone with enough GPU memory to run them.

The architecture is where things get genuinely interesting. Qwen3.8 doesn't just scale up the standard transformer — it introduces a hybrid layout that interleaves Gated DeltaNet (a linear attention mechanism) with traditional Gated Attention across 92 layers. The pattern is structured as 23 blocks of four layers each, where three layers use Gated DeltaNet followed by one layer of full attention. This design lets the model maintain a 262K token context window natively, extensible up to 1 million tokens.

Key architectural specifications:

• 2.4T total parameters, 95B activated per token (MoE with 512 experts, 10 routed + 1 shared)
• 92 layers in a 23 × 4 structured layout mixing Gated DeltaNet and Gated Attention
• 248,320 token vocabulary with Multi-Token Prediction (MTP) training
• 262,144 token native context, extensible to 1,010,000 tokens
• Compatible with vLLM, SGLang, and TokenSpeed for inference

Benchmark Performance: Competing With Closed Models

The benchmarks are where Qwen3.8 makes its case. On PaperBench, a benchmark for research-paper comprehension and reproduction, Qwen3.8-Max scores 93.0 — beating Claude Opus 4.8 (80.3), Fable 5 (88.8), and GPT-5.6 Sol (90.5). On Terminal Bench 2.1, which measures coding agent performance in real terminal environments, it scores 86.6, second only to GPT-5.6 Sol's 88.8.

Other notable results:

• GPQA Diamond: 92.6 (competitive with Claude Opus 4.8 at 92.0 and GPT-5.6 Sol at 94.1)
• HealthBench: 60.2, beating all competitors including GPT-5.6 Sol (55.3)
• IFBench (instruction following): 82.8, significantly ahead of all frontier models
• PLawBench (legal reasoning): 73.2, ahead of GPT-5.6 Sol (72.3) and Fable 5 (70.2)
• WideSearch: 81.9, nearly matching Fable 5's 81.2 and well ahead of Qwen3.7-Max's 75.2

Where Qwen3.8 falls short is on some coding agent benchmarks — SWE-bench Pro (67.7) trails Fable 5 (80.0), and DeepSWE 1.1 (56.6) is behind both Fable 5 (70.0) and GPT-5.6 Sol (73.0). But for general-purpose agentic tasks, research, and instruction following, it's in the same league as models that cost billions to develop and are locked behind paid APIs.

The Mixture-of-Experts Efficiency Story

The 2.4T-to-95B ratio is the real engineering triumph here. During inference, only about 4% of the model's total parameters are activated for any given token. The routing system selects 10 of 512 expert networks plus 1 shared expert, meaning the computational cost is closer to a 95B dense model than a 2.4T one. This is how you build a model that has more capacity than GPT-5 but runs on a fraction of the compute.

For context, a 95B dense model can run on a single 8×H100 node with reasonable batch sizes. The full 2.4T model needs to be distributed across multiple nodes — but the inference cost per token is determined by the activated parameters, not the total. This is the same principle that powers DeepSeek's models and Google's Gemini, but at a scale we haven't seen in open weights before.

The expert routing also means the model can specialize. With 512 experts, different knowledge domains and skills can be distributed across different expert networks, reducing interference between tasks. The Gated DeltaNet layers add another efficiency layer — linear attention mechanisms have lower memory and compute costs than full self-attention, which is what enables the 262K native context window without prohibitive KV cache requirements.

Agent Capabilities and Practical Use

Qwen3.8 was designed with agentic tasks in mind — not just question-answering, but multi-step task execution where the model needs to plan, use tools, handle environment feedback, and carry tasks through to completion. The benchmarks reflect this:

• Toolathlon Verified (Pass@1): 72.5 — close to Fable 5 (77.9) and ahead of Qwen3.7-Max (49.7)
• CoWorkBench: 74.8, nearly matching Fable 5's 75.9
• WorkSpaceBench: 67.7, slightly ahead of Fable 5 (68.7 is the leader)
• AndroidBench: 75.1, second only to Fable 5 (84.5)

The model also introduces flexible thinking controls. The reasoning_effort parameter lets you tune how much the model thinks before responding — useful for balancing latency and quality. The preserve_thinking feature retains reasoning context across conversation turns, which is critical for multi-step agent workflows where the model needs to remember why it made earlier decisions.

How to Run Qwen3.8

The model weights are available on Hugging Face under the Qwen/Qwen3.8-2.4T-A95B repository. For managed inference, Qwen Cloud offers the Qwen3.8-Max variant with additional features including vision input, non-thinking mode, 1M context by default, and built-in tool support.

For self-hosting, the model is compatible with:

• vLLM — the most popular high-throughput inference engine for LLMs
• SGLang — optimized for structured generation and agentic workloads
• TokenSpeed — for high-performance token streaming

Given the 2.4T total parameter count, you'll need significant VRAM just to load the weights — roughly 4.8TB in FP16 or 2.4TB in FP8. But remember, the activated parameters are only 95B, so inference throughput is similar to running a 70B-class dense model once the weights are loaded across your GPU cluster.

What This Means for the AI Industry

Qwen3.8's release continues a trend that's reshaping the AI landscape: open-weights models closing the gap with proprietary frontier models. When a free, downloadable model can beat GPT-5.6 Sol on HealthBench, PLawBench, and IFBench, the value proposition of paying per-token for API access to closed models becomes harder to justify for many use cases.

The competitive dynamics are clear:

• Open models are catching up on general capabilities while maintaining advantages in cost, privacy, and customization
• Proprietary models still lead on specialized coding agent tasks (SWE-bench, DeepSWE) — but the gap is narrowing each generation
• The MoE architecture is proving that raw parameter count matters less than expert routing efficiency — 95B activated can match 500B+ dense models
• Alibaba is positioning Qwen as the open-weights alternative to Anthropic and OpenAI, targeting developers who need frontier-quality AI without vendor lock-in

The Hacker News discussion around this release (627 points and climbing) highlights the community's excitement — but also raises practical questions about who can actually run a 2.4T model. The answer today is: cloud providers, well-funded labs, and anyone willing to rent serious GPU infrastructure. But the 95B activation footprint means inference costs are manageable once you've made the upfront investment in loading the model.

For developers building AI applications, Qwen3.8 represents a new option in the growing menu of frontier-class open models. Whether you're building coding agents, research tools, or general-purpose assistants, having a 2.4T MoE model with open weights that competes with the best closed models — and that you can fine-tune, inspect, and deploy on your own infrastructure — is a meaningful step toward democratizing access to top-tier AI capabilities.