Flint
39.8B parameter open-weight model. Sparse mixture-of-experts coding model from Recursive: 39.8B total parameters (Hugging Face safetensors total) with about 2.5B active per token, 49 layers (2 dense, then 47 MoE layers of 256 routed experts with 6 selected plus a shared expert), 35 sliding-window and 14 full-attention layers, 4 KV heads, and a 262,144-token context (config.json and model card read 2026-10-23). Built for coding agents: reasoning, tool calling and multi-turn tool use. Apache 2.0. The flint tag was verified against the Ollama registry manifest (checked 2026-10-23): it serves a single 22.86 GB NVFP4 file, not a Q4_K_M, and it needs Ollama 0.40.9 or newer, so an older Ollama refuses it. This project measured it on an RTX 3080 (10 GB) under Ollama 0.40.10 with nothing else on the card: about 8.8 GB on the GPU and the remaining expert weights mapped in system RAM, 46 to 47 tok/s once warm at both a 4,096 and a 32,768 context, 30.9 tok/s on the cold first run, and a correct answer on a short coding task (2026-10-23, docs/lab/radar-2026-10-23-model-validation.md). The NVFP4 figure is the full weight footprint, not a VRAM requirement: plan around 24 GB of combined VRAM and free system RAM, and expect the CPU-side experts, not the GPU, to set the speed.
Recursive · Flint
Model overview
Flint is a sparse mixture-of-experts model for coding agents: 39.8B parameters in total, but only about 2.5B work on each token, which is why it stays usable when most of the model lives in system RAM. It investigates codebases, fixes bugs, writes tests and calls tools, with a 262K context window. The Ollama download is a 22.9 GB NVFP4 file, so on a 10 to 16 GB graphics card it runs split between VRAM and RAM; this project measured about 47 tokens a second that way on a 10 GB card. You need Ollama 0.40.9 or newer to run it.
Editorial review
VRAM figures are empirical estimates. Actual usage varies by runtime, context length, and system configuration. Verify on your specific hardware before production use.
Will Flint run on your machine?
Flint is 39.8B parameters and needs 24.4 GB of VRAM at NVFP4: 22.9 GB of weights plus 1.5 GB of runtime overhead for the inference server itself.
VRAM by quantization
| Quantization | Weights | Needs (with overhead) | Quality |
|---|---|---|---|
| NVFP4 | 22.9 GB | 24.4 GB | good |
Fit on common hardware at NVFP4
| Hardware | Memory the model can use | System RAM | Verdict |
|---|---|---|---|
| CPU Only | None (CPU only) | 16 GB | Too large |
| RTX 4060 Laptop | 8 GB | 16 GB | Too large |
| RTX 3060 (12GB) | 12 GB | 32 GB | Too large |
| RTX 4060 Ti (16GB) | 16 GB | 32 GB | Too large |
| RTX 3090 | 24 GB | 64 GB | CPU offload |
| Apple Silicon (Unified Memory) 36 GB | 27 GB of 36 GB | 36 GB | Comfortable |
| RTX 5090 | 32 GB | 64 GB | Comfortable |
| Apple Silicon (Unified Memory) 48 GB | 36 GB of 48 GB | 48 GB | Comfortable |
Comfortable means VRAM clears the requirement by 2 GB or more. Tight means it covers the requirement with no margin. CPU offload means the model does not fit in VRAM but system RAM is at least 1.6× the weights, so it will run at reduced speed: expect roughly 1–5 tokens per second. Figures are weights plus a fixed runtime overhead and exclude KV-cache growth, which scales with context length.
Apple Silicon shares one pool of memory between the system and the GPU, so a model cannot use all of it. These rows apply the same 75% usable fraction the Compatibility Checker uses, which is why a 36 GB Mac is graded on less than 36 GB.
Want a broader hardware plan around Flint? Open the PC Builder for the whole 30B / 32B class → The Builder plans across that class; it does not replace the exact-model fit verdict above.
Ready to run this model locally?
Find a compatible interface in our Local AI Tools directory →