Qwen3.8 Open Weights Bring Alibaba Close to Claude, With Caveats

What Alibaba released

Alibaba has published the weights of its flagship model, Qwen3.8-Max. The release is a major step for open-weight AI, but a closer look at the license, the benchmark claims and the hardware needed to run it complicates the claim that it is simply free. The company released the model as Qwen3.8-2.4T-A95B on August 12, 2026, about nine days after the hosted Max version went generally available on August 3.114 At 2.4 trillion parameters, it was the second-largest and second-strongest open-weights model released by mid-August, behind Moonshot's Kimi K3.11

The dates differ slightly between outlets. One puts the hosted API launch on August 2 and the weight release on August 13.5 Most others give August 3 and August 12.111617 That gap is minor. The licensing disagreement, covered below, matters more.

Two days after the large model, Alibaba released Qwen3.8-27B. This smaller dense model includes features that were left out of the big open release: image input and a non-thinking mode.11 For most developers, that 27B model may matter more than the giant one.

How close it gets to Claude and ChatGPT

The claim that Qwen nearly matches the best closed models depends on which numbers you trust. Alibaba's own scores for Qwen3.8-Max include 92.6 on GPQA Diamond, 67.7 on SWE-bench Pro, 86.1 on OSWorld-Verified and 43.6 on Humanity's Last Exam. The company compares the model with GPT-5.5, Claude Opus 4.7 and Gemini 3.1 Pro.4 Alibaba has also said the model trails only Anthropic's Claude Fable 5 among current frontier systems.1

Outside observers are more cautious. One review says Alibaba has not published an official benchmark table for Qwen3.8-Max. It found only one independent head-to-head result, an architecture test where the preview scored 80 out of 100 against Kimi K3's 83.1 A model-routing company pointed out that the 86.6% Terminal Bench figure came from Alibaba's own testing setup, so comparisons with rivals' self-reported numbers prove little.4 In practice, Qwen is close to the frontier but not at it, and much of the gap that remains is in exactly the coding and agent tasks where Anthropic leads. A comparison of free coding models makes the same point: Claude Fable 5 scores 95.0% on SWE-bench Verified, far ahead of the leading open-weight models, which sit around 77–81%.12

Not quite free: the license question

The coverage diverges most on licensing, and "free" depends on the answer. One outlet reported that both Qwen3.8-Max and Qwen3.8-27B ship under Apache 2.0.5 Several other accounts disagree. They say the 2.4T model uses a custom Qwen3.8-Max license, and that providers earning more than US$50 million a year must sign a commercial agreement with Alibaba. Only the 27B model gets the permissive Apache license.111617 One analysis also notes that the open flagship is a text-only version of the hosted product.17

The custom-license account is the more reliable one. More sources report it, and it fits a pattern that has built up over three years. Alibaba started in 2023 with a custom license that was free below 100 million monthly users. It moved most models to Apache 2.0 by 2024 and every open model by April 2025. It then kept its flagship Max models API-only from September 2025.16 Through the first half of 2026, Chinese AI companies were reported to be turning to proprietary models to make money, and Alibaba kept its Plus and Max tiers closed while releasing mid-size models openly.16 Releasing the August weights with a revenue threshold is a partial step back toward openness. Alibaba still keeps the right to charge the largest commercial users.

The real story is efficiency

The headline is about frontier capability, but the more important pattern in Alibaba's recent research is how much performance it gets from each unit of computing power. Qwen3.8-Max is a sparse mixture-of-experts (MoE) model: it splits its parameters into many "expert" sub-networks and uses only a few of them for each token. Only about 95 billion of its 2.4 trillion parameters are active for any one token. That design is the main reason anyone outside a hyperscaler can host it at all.5

The trend is even clearer in smaller models. Qwen3.8-Flash, released August 26, has 125 billion total parameters but uses only 6 billion per token. It spreads that capacity across 512 experts, routing each token to ten of them plus one shared expert.3 One report says this model scored 62.5 on SWE-bench Pro, beating DeepSeek V4 Pro's 55.4 even though DeepSeek's model uses 49 billion active parameters.3 Alibaba also says the Flash-Next design needed about one-ninth of the training compute of Qwen3.7-Plus.3 Another report puts the saving at roughly 90%.9

One analysis highlights a less obvious technique. In addition to the MoE core, Flash-Next stores a 51-billion-parameter table of n-gram embeddings, meaning saved representations of short word sequences. Looking those up is far cheaper than computing them.7 On that analysis's figures, the model uses about 6 billion of 176 billion total parameters, a ratio near 29 to 1, compared with about 25 to 1 for the Max flagship.7 If that holds, Alibaba is getting part of its efficiency by replacing computation with memory lookups, not just by routing tokens more cleverly.

This work builds on earlier architecture changes. Qwen3.5, released in February, combined Gated Delta Networks with sparse MoE. Gated Delta Networks are a form of linear attention that keeps processing costs from growing as fast as the input gets longer.2 Its open 397B model uses only 17 billion parameters per token.16 Alibaba says the hybrid design decodes text 8 to 19 times faster than the earlier Qwen3-Max generation at about 60% lower cost.6 One developer guide argues that the linear-attention memory explains why the model does well on long, multi-step agent tasks.6 That is an interpretation, not a measured cause, but it is a reasonable one.

What runs on ordinary hardware

For developers and researchers, efficiency only counts if the model runs on hardware they can get. Qwen3.5-35B-A3B uses 3 billion of its 35 billion parameters per token and fits in about 20GB of memory once quantized. One test found it beats the much larger Qwen3-235B on MMLU while running about three times faster on a Mac.8 Qwen3.8-27B looks like the best value of the August releases. One benchmark table shows it scoring 61.7 on SWE-bench Pro, above the 53.4 listed for Claude Opus 4.6 Max. On Artificial Analysis's independent Intelligence Index it scores 33.70, slightly above that Claude model's 31.95.6

The flagship is another matter. The full-precision weights take about 4.9TB and need a multi-GPU cluster. Hosted versions on OpenRouter and similar services cost about $2 per million input tokens and $6 per million output tokens.16 One outlet notes that downloadable weights do nothing to remove the GPU infrastructure and engineering skill needed to serve a model this size at usable speed.5 For most users, "free" in practice means free to run the 27B model, or cheap to call the hosted versions.

Why it matters

The broader effect is that Qwen is becoming the default open base model. When developers fine-tune an open model, it is very often a Qwen model.17 Alibaba had released more than 400 open Qwen models, downloaded over a billion times, by March.16 Releasing flagship weights strengthens that position even with the revenue limit attached.

Alibaba's other recent announcements raise the stakes. The company has confirmed that Qwen 4 is in training and has unveiled the Zhenwu V900 accelerator chip, which has 216GB of on-package memory.7 It also says Qwen3.8-Max ran 33 rounds of automated self-optimization over about a month, lifting its Artificial Analysis score from 40 to 45.7 That claim has not been independently checked and should be treated with caution.

Alibaba's open models do not beat Claude or ChatGPT on hard agent work, and the free release comes with commercial conditions. What it has shown is that frontier-adjacent performance can be reached with a small fraction of the active compute. That matters more to open-model developers than any single benchmark result.