What was announced

Throughout August, NVIDIA is running a publication series highlighting local AI. In the entry dated 14 August 2026, the company announced day-zero support on RTX GPUs for the open model Qwen3.8-27B.

The model is positioned as the local companion to Qwen3.8-Max. This 27-billion-parameter open model is sized for a single GPU and built for responsive coding workflows with local files, tools and project context. The benefit NVIDIA emphasises is straightforward: developers can keep sensitive code, proprietary data and project context on their own systems while using capable AI assistance throughout the development process.

Performance and support

Two points stand out on the technical side. The first is coverage: day-zero support applies to RTX GPUs including NVIDIA RTX PRO cards and the GeForce RTX 5090. The second is speed: optimised with multi-token prediction, the model reaches 131 tokens per second on a single GeForce RTX 5090 with MTP enabled. That means fast local inference for agentic coding, tool use and longer development tasks.

The announcement also lists the tools where the model can be used immediately:

  • llama.cpp — a low-level, portable inference engine,
  • Ollama — the widely used tool that packages running models locally,
  • Unsloth — a library providing acceleration on the fine-tuning side,
  • LM Studio Bionic — a desktop environment for running local models.

All four support Qwen3.8-27B on NVIDIA RTX GPUs. NVIDIA presents this as the combination of RTX performance with the CUDA software ecosystem.

Why it matters

The real issue here is the sizing choice. Most coding agents today rely on very large models running in the cloud, which means code and project context leave the machine. A 27-billion-parameter model that runs at a meaningful speed on a single consumer-class card changes that balance for developers outside large organisations too.

The 131 tokens per second figure is symbolic in that respect. A coding agent does not just produce one response; it takes many steps, calling tools and reading results, so generation speed translates directly into waiting time. At that level of speed a local agent becomes comparable to a cloud-based alternative in everyday use.

Caveats

Even so, the picture should not be read as better than it is. The performance figure comes from NVIDIA's own announcement and has not been confirmed by independent benchmarks. Tokens per second also varies markedly with the quantisation level used, the context length and the shape of the workload; a single number does not represent every scenario.

A similar caveat applies on model quality: a 27-billion-parameter model is not expected to match the largest cloud models on every task. NVIDIA's own wording describes it as a local companion to Qwen3.8-Max rather than a replacement for it. The decision comes down to how much weight the wish to keep sensitive code in-house carries.