DeepSeek reads long prompts about 5% faster on AMD Strix / Gorgon Halo, and three fixes.
- Strix / Gorgon Halo: DeepSeek reads a 16K prompt about 5% faster (the prompt is read in 4K-token pieces on these GPUs, with the same answers).
- Long agent sessions: a request whose messages carry their content as parts no longer fails with an error (#717).
- Tool calls: a tool call with an unclosed parameter tag is handled as a malformed call instead of failing the request (#706).
- DeepSeek on Windows: the in-place memory trim settings given in a config now reach the engine (#728).
Maya is free; if it helps you, you can sponsor it: https://github.com/sponsors/mw00