A better and faster DeepSeek-V4.1-Flash Maya-XS, smaller downloads for unsloth's DeepSeek quants, and a cleaner Monitor.
- DeepSeek Maya-XS v2: about 12% closer to the original model (KLD 0.251 against 0.284) and about 20% faster at writing, at almost the same size (117 GB). The weights every token uses are now stored more compactly, which leaves room for more experts on the GPU, and 12 layers get a higher-quality format. Running the setup again asks whether to update an existing Maya-XS or keep it: the update downloads only the new model files (117 GB), keeps the n-gram tables, and removes the old files once the new ones are complete.
- One set of n-gram tables for every DeepSeek quant: unsloth's quants now use Maya's Q6_K n-gram tables, shared with Maya's own quants on disk. Measured with the same scorer, they lose nothing against the larger MXFP8 tables, and the Q8 download drops from 508 to 467 GB.
- Monitor: the peak decode speed sits small beside the live Decode figure, Decode and Prefill sit side by side on a phone, and Speculation shows the share of drafts kept as the main figure, with the drafter below it.
- Load and unload models from any API client or command: the
/models/load,/models/unloadand/models/sseendpoints are now on by default ("model_management": falseturns them off). --engine mayafor the server (the old name still works), and Maya introduces itself as Maya to MCP servers.
Maya is free; if it helps you, you can sponsor it: https://github.com/sponsors/mw00