Alibaba's Qwen3.8-27B Fits a Frontier Model on One GPU

Alibaba's Tongyi Lab released Qwen3.8-27B on Aug. 14, a 27.8 billion-parameter dense model that runs on a single 80-gigabyte-class GPU and scores even with DeepSeek V4 Flash, a rival open-weight model that needs two, independent testing firm Yotta Labs said.

The model is described in the QwenLM/Qwen3.8 repository on GitHub, published by the Qwen team at Alibaba. It ships under an Apache 2.0 licence, the same permissive terms Alibaba used for earlier Qwen releases, and the weights are posted on Hugging Face and ModelScope for direct download.

A dense 27.8 billion parameters, not a mixture of experts

Qwen3.8-27B is a dense model, meaning every parameter activates on every token, unlike a mixture-of-experts design that routes each token to a subset of the network. Two days before it shipped, on Aug. 12, Alibaba released a second model in the same family, Qwen3.8-2.4T-A95B, a mixture-of-experts flagship the GitHub repository describes only by its parameter notation without stating the total and active counts in plain language.

Qwen3.8-27B is natively multimodal, taking images and video as input alongside text, and supports a 262,144-token context window that Alibaba says can be extended to 1 million tokens. That is large enough to hold roughly 200,000 words of source material in a single request.

How it scores against DeepSeek V4 Flash and Kimi K3

Yotta Labs, a third-party infrastructure analysis firm, tested Qwen3.8-27B and put it at 52 on Artificial Analysis's Intelligence Index, an independently run benchmark aggregator. DeepSeek V4 Flash scores the same 52, Yotta Labs said, but needs two Nvidia H200 GPUs to run at full precision and lacks Qwen3.8-27B's image and video understanding.

Model	Artificial Analysis Intelligence Index	Hardware for full precision
Qwen3.8-27B	52	1 GPU, 56 GB VRAM (BF16)
DeepSeek V4 Flash	52	2 Nvidia H200 GPUs

Yotta Labs also compared the release to Kimi K3, a much larger open-weight model that it said ships as 1.56 terabytes of weights requiring cluster-scale hardware, a size class Qwen3.8-27B avoids entirely.

What it takes to run, from a data-centre GPU to a laptop

At BF16 precision, Qwen3.8-27B needs roughly 56 GB of VRAM, putting it on GPUs like the H100, H200 or RTX Pro 6000, Yotta Labs said. Quantized to FP8, that drops to about 28 GB, fitting a single L40S card. At 4-bit quantization, it needs 14 to 16 GB, within reach of a consumer RTX 4090. Yotta Labs cautioned that these figures cover the weights alone; the key-value cache used during inference adds more memory on top and grows with context length and how many requests run at once.

What Alibaba has not independently verified

Alibaba's own Hugging Face model card lists benchmark scores including 89.2 on GPQA Diamond, a graduate-level science reasoning test, and 61.7 on SWE-bench Pro, a software-engineering benchmark. The card does not state whether an outside evaluator ran these numbers, and Alibaba has not published a separate technical report addressing that gap. The Artificial Analysis Intelligence Index score Yotta Labs cited is the only figure in wide circulation that comes from a benchmark operator independent of Alibaba.

Qwen3.8-27B had drawn 15,726 likes and more than 7.3 million downloads in the past month on Hugging Face as of Sept. 19, according to the platform's own model page, making it one of the most-adopted open-weight releases of the past six weeks. Whether Artificial Analysis scores the larger Qwen3.8-2.4T-A95B model, and how it compares to closed frontier systems, is the next number to watch.

Similar efficiency claims have run into trouble before: Cactus Needle 3, an 8-to-29-megabyte automation model that claimed DeepSeek-level tool calls, drew scrutiny for benchmarks run only by its own developer. DeepSeek's own retirement of its V4 Pro tier in favor of a flash model, covered here in September, shows how quickly a vendor's efficiency pitch can reshuffle a product line.