Ollama 0.40.8-rc0 adds Apple MLX Flint model support

Brief written by the FOSSWire AI editor from Ollama How we write

October 21, 2026 1 min read

Key Takeaway

Ollama's first release candidate for v0.40.8 introduces support for Flint, a new model format on Apple's MLX framework that preserves numeric types and schema ordering.

Summary

Ollama v0.40.8-rc0 adds initial support for Flint models on Apple's MLX backend. The release preserves source JSON metadata alongside decoded API values, assembles ModelOpt expert banks during safetensors import, and adds optional learned sinks and device causal masks to shared MLX attention code. Existing models and API behavior remain unchanged.

What Flint support changes in MLX

Source JSON preservation: Flint retains original numeric types and schema order by keeping source tool JSON alongside decoded API values; existing API marshaling is unaffected
Safetensors import path: ModelOpt expert banks are now assembled during import; direct Flint NVFP4 creation is blocked before conversion
MLX attention extensions: Optional learned sinks and device causal masks added to shared attention implementation; existing callers keep no-sink/host-mask defaults, Flint opts in
Backward compatibility: Other models continue using their existing import and quantization paths unchanged

Original reference: github.com →