Qwen3.8-27B: A Dense Vision Model's Local Budget
MODELS & COSTS
Qwen3.8-27B: A Dense Vision Model's Local Budget
A 27B dense vision model presents a more manageable storage question than Qwen's largest MoE. It still needs a runtime, vision components and enough working memory for the chosen context. Apache-2.0 weights make inspection possible; the local performance remains a measurement to make.
In this article
You're getting 27 billion dense parameters, native image and video understanding, 262K context that stretches to a million, and benchmark jumps over the last Qwen that aren't small. DeepSWE 1.1 went from 13.3 to 42.2. Computer use went from 63.9 to 84.3. All of it is vendor-reported, and the weights are hours old, so nobody has independently confirmed a single number yet.
Architecture, inputs and license
Twenty-seven billion dense parameters means all 27 billion fire on every token, no sparse experts. On top of that it's a native vision-language model, so it reads images and video, not just text. The architecture is the hybrid Qwen3.5 foundation: a mix of Gated DeltaNet linear-attention blocks and gated attention blocks, with multi-token prediction on top.
Before any benchmark, the license is the thing that stands out. Plain Apache-2.0, no custom restrictions, no revenue caps, nothing you'd need a lawyer for. It's also ungated. The weights are downloadable right now, eighteen safetensors shards, no approval form.
The runtimes Qwen lists are Transformers, vLLM, SGLang, and TokenSpeed. Note what's not on that list yet: llama.cpp. More on that in a minute.
The jump from Qwen3.6-27B
Qwen3.6-27B was already one of the local crowd's favorite coding models. The new one jumps it across the board. DeepSWE 1.1 going from 13.3 to 42.2 is the headline, roughly a threefold jump on a benchmark that measures real multi-step software engineering. QwenSWEBench went from 49.3 to 79.0. LiveCodeBench v6 lands at 90.3.
These are Qwen's own numbers with their own harnesses, so read them as direction rather than gospel. The direction is consistent though: everything moved up, and a couple of things moved up a lot.
Vision and computer-use evaluations
This is what the "intelligence density" tagline is pointing at. Qwen3.6-27B basically couldn't see, 28.9 on BabyVision, and its computer-use score was 63.9. The new model jumps those to 65.7 and 84.3. On OSWorld, that 84.3 from a 27B dense model beats the Opus 4.6 Max column in Qwen's own table, which sits at 72.7. AndroidWorld is the same story, 81.9 versus 62.0.
Translation: a model you can run on one consumer GPU is now doing computer use, browser use, and mobile use at flagship-ish levels. That used to be the thing that required a 2.4T monster. It's the density pitch made concrete, and for once the tagline has a number behind it.
Local runtime and memory requirements
Twenty-seven billion dense sits in the local sweet spot. At 4-bit you're looking at roughly 17GB of VRAM, Unsloth's early estimate, which fits a 24GB card comfortably. 8-bit needs about 28GB and fp16 about 55GB. Add a little for the vision encoder on top.
The catch is the runtime. The hybrid DeltaNet architecture means the usual llama.cpp and GGUF route isn't guaranteed. If you're an Ollama person, you may be waiting for someone to land GGUF support first. If you've got a 24GB GPU and can run vLLM, you're in business today.
The DeepSWE and computer-use gains are vendor launch results. Independent reproduction should keep the evaluation harness fixed. Local usability also needs the selected quantization and runtime to support the hybrid architecture and its vision components.
Apache-2.0, downloadable weights and multimodal input make a concrete package to evaluate. They do not establish that this is the best coder or agent at its size. Compare the tasks that matter to the deployment, including failures and memory use.
A useful follow-up would report the exact artifact, working context length, peak memory and task completion on a supported runtime. Those details would help a local user more than predicting which model will become popular.
Comments
Post a Comment