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.
Core Facts
Timeline (last 90 days)
Flint's sparse architecture usually refers to a Mixture of Experts (MoE) architecture, where the total parameters determine model capacity and the activated parameters determine the compute per inference.
Claude Code's system prompts are tuned for Claude models, so using them directly with Flint may cause behavioral deviations, and an intermediate layer for request/response conversion and tool-call validation and retries is usually needed.
The benchmark data published by Recursive needs to be interpreted with caution, as benchmarks usually target specific task sets and differ from real-world development scenarios.
Evaluation results for open-source models are affected by quantization precision, prompt templates, and inference backends, and scores may fluctuate considerably under different configurations.
When using Flint with Claude Code, the performance bottleneck is often not the model itself but the reliability of tool calling.
Flint on Ollama suits developers who have data privacy requirements or want to work in an offline environment.
As an open-source model, Flint's adherence to tool-calling formats and system prompts varies, working for simple tasks but potentially producing format errors or tool-call failures on complex multi-step agent tasks.
By modifying environment variables or configuration files to set the Anthropic-compatible API base URL to point to the Ollama endpoint and specifying the model name as Flint, Claude Code's requests can be forwarded to a local Ollama instance.
To run Flint on Ollama, it is recommended to have at least 24GB of unified memory or VRAM, otherwise CPU offloading may be triggered and cause slower speeds.
Ollama's model library already supports Flint.
35 more timeline events
All Facts (20)
A 70B-parameter FP16 model requires approximately 140 GB of video memory, approximately 70 GB after INT8 quantization, and approximately 35–40 GB after INT4 quantization.
90%Verified量化会不可避免地引入精度损失，在逻辑推理等对准确性敏感的任务上可能出现性能下降
65%UnverifiedTo run Flint on Ollama, it is recommended to have at least 24GB of unified memory or VRAM, otherwise CPU offloading may be triggered and cause slower speeds.
50%UnverifiedFlint on Ollama suits developers who have data privacy requirements or want to work in an offline environment.
50%UnverifiedFlint's sparse architecture usually refers to a Mixture of Experts (MoE) architecture, where the total parameters determine model capacity and the activated parameters determine the compute per inference.
50%UnverifiedBy modifying environment variables or configuration files to set the Anthropic-compatible API base URL to point to the Ollama endpoint and specifying the model name as Flint, Claude Code's requests can be forwarded to a local Ollama instance.
50%UnverifiedThe benchmark data published by Recursive needs to be interpreted with caution, as benchmarks usually target specific task sets and differ from real-world development scenarios.
50%UnverifiedClaude Code's system prompts are tuned for Claude models, so using them directly with Flint may cause behavioral deviations, and an intermediate layer for request/response conversion and tool-call validation and retries is usually needed.
50%UnverifiedWhen using Flint with Claude Code, the performance bottleneck is often not the model itself but the reliability of tool calling.
50%UnverifiedEvaluation results for open-source models are affected by quantization precision, prompt templates, and inference backends, and scores may fluctuate considerably under different configurations.
50%UnverifiedOllama's model library already supports Flint.
50%UnverifiedAs an open-source model, Flint's adherence to tool-calling formats and system prompts varies, working for simple tasks but potentially producing format errors or tool-call failures on complex multi-step agent tasks.
50%UnverifiedFlint approaches the performance of Qwen 3.8 27B while being 3-6 times faster, and is the best performer among comparable sub-50B MoE models on benchmarks such as SWE Bench Pro, VR Verified, and Terminal Bench 2.1.
50%UnverifiedUnder 3-bit quantization and short-context scenarios, Flint can run on a single consumer-grade GPU with 24GB of VRAM.
50%UnverifiedFlint scores 89.7 on SWE-bench Pro v2, while Qwen 3.6 scores 73.8.
50%UnverifiedFlint is a 40B MoE model for coding agents released by Recursive, with 2.5B active parameters, supporting reasoning, tool calling, and a 256K context.
50%UnverifiedThe original commit message is "Close the thinking section when think:false so Flint answers directly".
50%UnverifiedFor default settings and prompts with think:true, Flint's behavior remains unchanged.
50%UnverifiedFlint is one of the models in the Ollama ecosystem and supports optional output of the thinking process.
50%UnverifiedFlint's Apache 2.0 license means it can be freely used in commercial scenarios and redistributed.
50%