Open weights
FreeApache 2.0 licenseDownload the weights from Hugging Face and run or adapt them on your own hardware. Official source 1 ↗ · Official source 2 ↗
Reviewed tool profile
Recursive Flint
Open-weight coding model from Recursive that runs fast on a laptop for local agentic coding.
Sources checked
- Pricing
- The model weights are free to download and use under the Apache 2.0 license, and Flint-40B can be picked for free from OpenCode's free tier. Cloud access is also offered through OpenRouter, but Recursive's pages do not list a price for it. Recursive's own Flint API is limited to select partners on request, with no public price. Official source 1 ↗ · Official source 2 ↗ · Official source 3 ↗ · Official source 4 ↗
- Available as
- Local on macOS with rec, Recursive's Mac inference engine, which needs an M4 Pro, M5 Pro or M5 Max with at least 48 GB of unified memory and about 30 GB of free disk space. · Local through Ollama, LM Studio Bionic, Unsloth and llama.cpp (Apple Silicon, or NVIDIA GPUs on Linux or WSL2), plus Hugging Face Transformers in Python and GPU serving with SGLang or vLLM. · Cloud through OpenCode's free tier and OpenRouter; a direct Recursive API is for select partners. Official source 1 ↗ Official source 1 ↗ · Official source 2 ↗ · Official source 3 ↗ Official source 1 ↗
- Primary group
- Coding & Development
Also consider
See all 6 similar tools ↓- OpenPilotOpen-source Windows desktop AI agent that edits files, runs commands and researches the web with your own OpenAI-compatible model.Free
- ColdteaAn agentic development environment spanning code, release testing, and monitoring.Free plan
- REAOpen-source MCP server and CLI that lets your coding agent reverse engineer apps, binaries and websites.Free
What it does
Flint is an open-weight coding model from Recursive (Recursive Superintelligence), built for fast agentic coding on local hardware such as a recent MacBook Pro. It is a mixture-of-experts model with 39.8 billion total parameters and about 2.5 billion active per token, designed to investigate codebases, fix bugs, implement features and write tests. Recursive says its small team built it from scratch in under four months with its automated research system, the Eureka Machine. The weights are on Hugging Face under Apache 2.0, and it can be run locally through Recursive's own rec engine, Ollama, LM Studio, llama.cpp and other runtimes, or used in the cloud through OpenCode's free tier and OpenRouter.
Official source 1 ↗ · Official source 2 ↗ · Official source 3 ↗Features
- Small active size for local speed. A mixture-of-experts model with 39.8B total parameters but only about 2.5B active per token, so it runs fast on local hardware. Official source 1 ↗ · Official source 2 ↗
- Agentic coding benchmarks. Recursive reports 54.9% on Terminal-Bench 2.1, 89.7% on SWE-bench Pro v2 and 74.2% on SWE-bench Verified, and says Flint leads MoE models under 50B total parameters on these benchmarks. Official source 1 ↗ · Official source 2 ↗
- Reasoning and tool use. Text in, text out, with always-on reasoning, multi-turn tool use and OpenAI-compatible tool calls. Official source 1 ↗ · Official source 2 ↗
- rec engine for Macs. Recursive's rec engine downloads the model, starts a local server and launches Pi or OpenCode in one command; Recursive reports 1.4–2.6× llama.cpp's token throughput on an M4 Pro and over 125 tokens per second on an M4 Pro MacBook Pro. Official source 1 ↗ · Official source 2 ↗ · Official source 3 ↗
- Works with common coding harnesses. Connects to Pi (recommended), OpenCode, Codex, VS Code, Cline and OpenHands, for example through Ollama's launcher. Official source 1 ↗ · Official source 2 ↗
- Long context. The documented configuration is a 262,144-token context with up to 65,536 output tokens; Recursive's announcement says Flint supports up to 512k tokens but recommends staying within 256k. Official source 1 ↗ · Official source 2 ↗
- Open weights in several formats. Apache 2.0 weights on Hugging Face with BF16 and NVFP4 checkpoints, an SGLang serving kit, and packaged quantized builds for Ollama and LM Studio. Official source 1 ↗ · Official source 2 ↗ · Official source 3 ↗ · Official source 4 ↗
Pricing & free limits
OpenCode free tier
FreeHosted, in OpenCodeChoose Flint-40B from the free tier in OpenCode's model picker; OpenCode can also connect through an OpenRouter account. Official source 1 ↗
Flint API (partners)
No public priceBy requestDirect Recursive API access is limited to select partners, who get an assigned endpoint, model ID and API key. Official source 1 ↗ · Official source 2 ↗
Billing details
Hosted API terms and data practices are separate from the Apache 2.0 weight license. Official source 1 ↗
Recursive's pages list OpenRouter as a cloud option but do not state an OpenRouter price. Official source 1 ↗
Free access & limits
The weights are free under Apache 2.0; running locally avoids API fees but needs capable hardware (for rec, a Mac with at least 48 GB of unified memory). Official source 1 ↗ · Official source 2 ↗
Flint-40B is available in OpenCode's free tier with no manual configuration. Official source 1 ↗
Best for & limitations
Daily local coding with private code: using Flint as a coding model on your own machine without sending data to a remote server. Official source 1 ↗
Agentic coding work: understanding repositories, generating code, making edits and running tests through a coding harness. Official source 1 ↗
Self-hosted serving on GPUs: running the BF16 checkpoint on one B300 or H200 GPU per replica with the SGLang kit. Official source 1 ↗
- Needs more guidance. Recursive says Flint may need more guidance than larger closed-source coding models; split complex tasks into smaller subtasks with explicit instructions. Official source 1 ↗
- Review its output. The model can produce incorrect or harmful content, faulty code or inappropriate tool calls, so test generated code and review changes. Official source 1 ↗
- Long reasoning. Long reasoning can add latency and use up the output budget before an answer is complete. Official source 1 ↗
- Hardware and runtime limits. rec only runs on M4 Pro, M5 Pro or M5 Max Macs with 48 GB or more and handles one generation at a time; Transformers modeling code is not included in the Hugging Face repo yet. Official source 1 ↗ · Official source 2 ↗
Similar tools
Tools like Recursive Flint
Published tools that share its categories.
OpenPilotOpen-source Windows desktop AI agent that edits files, runs commands and researches the web with your own OpenAI-compatible model.
Visit ColdteaAn agentic development environment spanning code, release testing, and monitoring.
Visit REAOpen-source MCP server and CLI that lets your coding agent reverse engineer apps, binaries and websites.
Visit 
Mouse-1A fast decision model that scores options for agent harnesses instead of generating text.
Visit 
RunInfraHosted open-model API for coding agents, with cheap cached input and monthly coding plans from $10.
Visit 
ExaWeb search API for AI agents, with page contents, deep-research agents, monitors and an MCP server.
Visit