Flint
A model that is sensitive to numeric types and schema ordering, supports the NVFP4 quantization format, and requires a specific conversion process to load weights correctly.
Timeline (last 90 days)
Recursive released Flint, an Apache 2.0-licensed coding model.
Flint's Apache 2.0 license means it can be freely used in commercial scenarios and redistributed.
Flint is positioned as a coding agent rather than a general-purpose chat model.
Flint achieves a generation speed of 125+ tokens per second on an M4 Pro MacBook Pro.
Flint's officially announced SWE-bench Verified score is 74.2%.
Flint has 40B total parameters and uses a MoE architecture, activating only 2.5B parameters during inference.
Flint is a free coding model that can run locally on a laptop.
Recursive is a startup that has released a coding model called Flint.
Most of Flint's development process was completed by Recursive's own AI research system.
Recursive views Flint as an early test of AI self-improvement.
12 more timeline events
All Facts (20)
Flint has 40B total parameters and uses a MoE architecture, activating only 2.5B parameters during inference.
50%UnverifiedFlint's Apache 2.0 license means it can be freely used in commercial scenarios and redistributed.
50%UnverifiedFlint is positioned as a coding agent rather than a general-purpose chat model.
50%UnverifiedFlint achieves a generation speed of 125+ tokens per second on an M4 Pro MacBook Pro.
50%UnverifiedFlint's officially announced SWE-bench Verified score is 74.2%.
50%UnverifiedFlint is a free coding model that can run locally on a laptop.
50%UnverifiedRecursive is a startup that has released a coding model called Flint.
50%UnverifiedMost of Flint's development process was completed by Recursive's own AI research system.
50%UnverifiedRecursive views Flint as an early test of AI self-improvement.
50%UnverifiedThe original article does not detail which specific parts of Flint's R&D the AI system participated in (such as architecture search, training hyperparameter tuning, data filtering, or code implementation).
50%UnverifiedThere are already many open-source options for local code models, and Flint faces real competition; its competitiveness depends on metrics such as code completion quality, multilingual support, and editor integration experience.
50%UnverifiedThe commit message does not disclose the source of the Flint model, its parameter count, or benchmark data.
50%UnverifiedThe v0.40.8 update makes targeted adaptations for Flint at three levels: data serialization, quantized weight import, and the attention mechanism.
50%UnverifiedDuring safetensors import, the MLX framework assembles the ModelOpt expert banks for Flint.
50%UnverifiedThe MLX framework refuses to directly create Flint's NVFP4 weights before conversion; Flint's NVFP4 weights require a specific conversion process rather than being instantiated directly at the import stage.
50%UnverifiedIn the prompt construction stage, this version preserves the source tool JSON alongside the decoded API values, so that Flint can retain numeric types and schema order in the prompt.
50%UnverifiedThe current information comes from a single commit description and lacks benchmark data and detailed specifications of the Flint model, so the actual performance gains remain to be verified by the community.
50%UnverifiedExisting attention callers still use the old path with no sinks and a host-side mask, and only Flint enables the new behavior.
50%UnverifiedThis version adds optional learned sinks and a device-side causal mask to the shared MLX attention mechanism.
50%UnverifiedThis version refuses to directly create Flint's NVFP4 model before conversion is complete.
50%