← Back to feed

Alibaba announced Qwen3.8-Max on August 3 as an API-only release, promised open weights “within about a week,” and then shipped eleven days later instead. That timeline matters less than what actually landed: both Qwen3.8-2.4T-A95B and Qwen3.8-27B on Hugging Face and ModelScope under Apache 2.0, with no revenue threshold, no application requirement, and no geofencing.

This is a pattern break for Alibaba. Every previous Qwen flagship generation held back the large model at launch and released smaller open-weights versions first, sometimes weeks later. Qwen3.8 ships both simultaneously.

The Two Models

Qwen3.8-2.4T-A95B is the Max-class model: 2.4 trillion total parameters with 95 billion active per token, matching the hosted Qwen3.8-Max API on specifications. Context window is 262,000 tokens by default, extendable to 1 million via YaRN. Modalities: text and image input. The open weights are BF16 safetensors compatible with vLLM and SGLang.

Alibaba’s vendor-reported benchmark table puts it at 86.1% on OSWorld-Verified (desktop computer use, ahead of Claude Fable 5 and GPT-5.6 Sol), 86.6% on Terminal-Bench 2.1, 93.0% on PaperBench, and 67.7% on SWE-bench Pro: strong on agentic computer use, behind Fable 5 by 12 points on repository-scale software engineering. These are self-reported numbers not yet independently reproduced on the computer-use benchmarks; the SWE-bench Pro figure similarly lacks a neutral-harness confirmation.

Qwen3.8-27B is the dense, self-hostable companion: 27 billion parameters, 55.6 GB of BF16 safetensors across 18 shards, 262K native context with 1M via YaRN. It processes text, images, and video, including multi-hour video content and document-heavy inputs. The model includes a flexible thinking mode enabled by default.

Hardware requirement: high-spec MacBook Pro or Mac Studio handles 27B at full precision. Consumer-grade GPU users are covered by community quantizations: 303 GGUF variants had appeared on Hugging Face by August 15. The 8-bit versions run on cards with 24 GB VRAM; 4-bit builds go lower.

Alibaba’s own head-to-head against Meta’s Muse Glimmer-30B puts Qwen3.8-27B ahead by more than 20 points on Terminal-Bench 2.1 (73.0 versus 51.7), with leads on IFBench and GPQA Diamond. These are vendor comparisons.

Why This Release Matters

The Max-class open-weight question has been the main unresolved item since DeepSeek V4 Pro shipped MIT-licensed weights in July and Kimi K3 pushed 2.8 trillion parameters to Hugging Face. Alibaba positioning Qwen3.8-Max as a closed API while DeepSeek and Moonshot released their flagship weights at roughly the same capability tier would have been a notable retreat.

Shipping both models simultaneously under Apache 2.0 closes that gap. It also signals something specific about Alibaba’s competitive theory: distribution through open weights matters enough to give up the access moat on a model that was still earning API revenue days earlier.

Apache 2.0 means no additional restrictions on commercial deployment, fine-tuning, or derivative model creation. The license cannot be revoked.

Current Inference Support

vLLM and SGLang both added Qwen3.8-27B support within days of the weight drop. The 2.4T model requires multi-node setup for full-precision serving. Community builds for the 27B include unsloth and lmstudio-community GGUF variants optimized for Ollama and LM Studio one-command deployment.

The hosted Qwen Cloud API for Qwen3.8-Max and Qwen3.8-27B runs $2.00 per million input tokens and $6.00 per million output tokens: with cached input as low as $0.17 per million. For users who want the 1M-token context window or managed inference without infrastructure overhead, the hosted path remains the practical route; vLLM with multi-node for the 2.4T is not a weekend project.

Benchmark Caveats

The headline numbers above come from Alibaba’s own August 3 launch post. OSWorld-Verified and Terminal-Bench 2.1 figures for Qwen3.8-Max have not been independently reproduced on the same harness versions at time of writing. The model’s competitive position on coding specifically: 67.7% SWE-bench Pro against Fable 5’s 80.0%: is a meaningful gap that OSWorld leadership does not offset for teams whose primary workload is repository-scale software work.

The practical read: strong for agentic computer use and document-heavy tasks at $2/$6 per million tokens, with open weights that allow self-hosting and fine-tuning. Not the top pick for SWE-bench-style coding benchmarks at current vendor-reported numbers.