Alibaba open-sourced Qwen3.8-27B on Aug. 14, releasing a 27-billion-parameter dense native multimodal model under the Apache 2.0 license. According to the input, the release carries no monthly active user or revenue threshold. The model supports a native 262K context window and can be extended to 1 million tokens through YaRN.
In the official model card, Alibaba listed 17 evaluations in total. Fourteen of them included side-by-side scores against Claude Opus 4.6 Max. Across those 14 comparisons, Qwen3.8-27B posted 9 wins and 5 losses. The figures all came from Alibaba’s own publication, and the input said there has been no third-party independent replication so far.
Most of the gains showed up in visual agent workloads
The strongest results were concentrated in tasks tied to visual understanding, interface interaction and instruction following.
- MathVision: 94.6 versus 65.5, a lead of 29.1 points
- CharXiv: 90.2 versus 66.0, a lead of 24.2 points
- AndroidWorld: 81.9 versus 62.0, a lead of 19.9 points
- IFBench: 79.5 versus 62.5, a lead of 17.0 points
- QwenSWEBench: 79.0 versus 63.8, a lead of 15.2 points
- OSWorld-Verified: 84.3 versus 72.7, a lead of 11.6 points
- SWE-bench Pro: 61.7 versus 53.4, a lead of 8.3 points
The input also noted that QwenSWEBench is Alibaba’s own benchmark, which limits how much weight should be placed on that result by itself.
The losses were all tied to reasoning and longer coding chains
The five weaker results shared a clear pattern. They were concentrated in pure text reasoning, code generation and terminal-style coding tasks.
- Terminus terminal coding: 73.0 versus 78.2, down 5.2 points
- NL2Repo-Bench code generation: 42.3 versus 47.6, down 5.3 points
- GPQA Diamond: 89.2 versus 91.3, down 2.1 points
- OmniDocBench: down 0.3 points
- RealWorldQA: down 1.0 point
Set side by side, the structure is straightforward: Qwen3.8-27B outperformed on tasks that involve reading screens, understanding charts and pressing buttons, while it trailed in pure reasoning and multi-step long-chain coding. That was how the input framed the model’s profile as well.
About 17 GB of VRAM can load the model, but that does not settle agent deployment
Hardware cost was one of the most concrete selling points in the release. The input said the original BF16 weights require about 56 GB of VRAM, implying an 80 GB-class card. FP8 lowers that to about 28 GB, which would call for a 48 GB card. At 4-bit quantization, memory use falls to roughly 14 GB to 17 GB, which means the weights can fit on a 24 GB RTX 4090, while a 32 GB RTX 5090 leaves more room.
That figure only covers model weights. It does not include KV cache, concurrent agent sessions, CUDA graphs, runtime buffers or multimodal components. The model’s native context window is 262K, and the input pointed out that memory demand changes materially once an agent starts packing an entire codebase and many rounds of tool output into context.
So the statement that a consumer GPU can run the model is tied to weight loading. Whether a single consumer card can run it comfortably as an agent depends on how long and how heavy the task is.
Alibaba chose Claude Opus 4.6 Max as the comparison target
The comparison set in the model card used Claude Opus 4.6 Max. The input added that Anthropic, in a risk report published on Aug. 14, disclosed an unreleased internal Model 2 that outperformed the publicly served Fable 5. In that report, the reference point for a capability jump was described as the move from Opus 4.6 to Mythos Preview.
That means the benchmark table published for Qwen3.8-27B was measured against Claude Opus 4.6 Max, not against the other Anthropic models mentioned in the input.
Model card adds a reasoning_effort control
Alibaba also introduced a reasoning_effort parameter with three settings: xhigh as the default, plus medium and low. Developers can use it to decide how long the model should spend thinking based on task difficulty. Another default-on option, preserve_thinking, keeps reasoning content in message history.
According to the input, controls of this kind have often been seen on pricing pages for closed API products. Here, they were written directly into the model card for an open-source release.
No independent replication yet
The key limitation around the benchmark claims is the source of the numbers. The 9 wins and 5 losses, the 19.9-point lead on AndroidWorld, the 29.1-point lead on MathVision, the 5.2-point deficit on Terminus and the 2.1-point deficit on GPQA Diamond all came from Alibaba’s official model card. The input did not provide any outside verification or independent rerun of those tests.
What is established from the input is that Alibaba released the model, published these benchmark figures and disclosed the memory range for a 4-bit version. Whether those results hold under third-party testing remains unanswered in the material provided.