Recursive Flint logo, a faceted three-sided stone shape in red and dark brown tones, illustrating a guide to the Flint coding model

Recursive Flint: Power-User Guide to Local Setup, Specs and Harnesses

by

Recursive Flint is an Apache 2.0 open-weight coding model with 39.8 billion total parameters and about 2.5 billion active per token. It runs locally through Recursive’s rec engine on M4 Pro and M5 Macs, or through Ollama, LM Studio and llama.cpp, and it pairs best with the Pi harness.

Last updated: November 10, 2026

Table of contents

What Recursive released

Recursive announced Flint in a blog post dated October 18, 2026 as a mixture-of-experts (MoE) model built for fast agentic coding on local hardware. The company says a small team working with its automated research system, the Eureka Machine, built the model from scratch in under four months.

The release has three parts. The weights are on Hugging Face under Apache 2.0. A first-party engine called rec runs Flint on recent Macs and connects it to a coding harness. And Recursive publishes setup guides for Ollama, LM Studio, Unsloth, llama.cpp, SGLang and vLLM.

Recursive positions Flint as a daily coding model for work you want to keep on your own machine, offline and without API fees. Its blog also says Flint may need more guidance than larger closed models, so break complex tasks into smaller steps.

Flint specifications

These figures come from the official model card:

Property Flint
Total parameters 39.8B (rounded designation 40B-A2.5B)
Active parameters per token About 2.5B, including the output head
Layers 49: 2 dense layers, then 47 MoE layers
Experts 256 routed experts per MoE layer, 6 selected per token, plus one shared expert
Attention 35 sliding-window layers (513 tokens) and 14 full-attention layers
Context configuration 262,144 tokens, prompt and output combined
Input and output Text only
Knowledge cutoff June 2026
License Apache 2.0

The BF16 checkpoint is about 80 GB. Quantized variants are published separately for Apple Silicon and NVIDIA runtimes, and Recursive says its NVFP4 model scores within 0.7 points of the 16-bit model across its benchmarks.

Benchmarks and how to read them

In Recursive’s release evaluations, Flint scores 54.9% on Terminal-Bench 2.1, 89.7% on SWE-bench Pro v2 and 74.2% on SWE-bench Verified, according to the model card. These are publisher-reported numbers. Terminal-Bench used the Terminus 2 harness and SWE-bench used mini-swe-agent, with average pass@1 over at least five attempts.

Recursive says Flint leads MoE models under 50 billion total parameters on those agentic coding tests and rivals dense models of 25 to 40 billion parameters while running more than five times faster. On speed, the blog reports more than 125 tokens per second on an M4 Pro MacBook Pro and more than 150 tokens per second on a 64 GB NVIDIA RTX Spark at short contexts.

For a comparison with another new open coding model, see our Ripple-65B Preview guide.

How to run Flint on a Mac with rec

Recursive’s rec documentation lists an M4 Pro, M5 Pro or M5 Max Mac with at least 48 GB of unified memory, about 30 GB of free disk space, and Node.js 22 with npm for the harnesses. rec checks your chip and memory and will not download the model on an unsupported Mac.

  1. Install rec with the script from the rec repository: curl -fsSL https://raw.githubusercontent.com/recursive-oss/rec/main/install.sh | bash. It installs to ~/.local/bin.
  2. Open a new terminal and run rec doctor to confirm your hardware and installation.
  3. From your project folder, run rec launch pi. The first run downloads the model and a matching Pi version, then starts the local server.
  4. Prefer OpenCode? Run rec launch opencode instead.
  5. To use your own client, run rec serve, wait for “flint is ready”, and connect to http://127.0.0.1:11500/v1 with model ID flint and the key from rec auth token. An Anthropic-compatible endpoint is also available at /v1/messages.
  6. When you finish, run rec stop. The server keeps running after you close the terminal.

Recursive recommends Pi as the harness and says informal reports suggest workflows roughly 1.5 times faster than OpenCode. Our Pi Coding Agent 1.5 guide covers Pi’s latest update. Note that rec is optimized for one session at a time.

Other ways to run Flint

  • Ollama. The Ollama library page lists flint:40b at 23 GB with a 256K context window, and ollama launch opencode --model flint or ollama launch claude --model flint connect a harness. Recursive’s Ollama guide uses the tag flint:40b-a2b-qat-int4. See our Ollama 0.40.15 guide for recent launcher changes.
  • LM Studio Bionic. Run lms get recursivesi/flint, then lms load --identifier flint and lms server start --port 1234, per Recursive’s LM Studio page. Our LM Studio Bionic guide covers the app.
  • llama.cpp, SGLang and vLLM. Use a compatible GGUF for llama.cpp. The SGLang kit in the model repository targets one B300 or H200 GPU per replica; background on the runtime is in our llama.cpp 0.9.0 guide.
  • Hosted options. OpenCode’s free tier offers Flint-40B, as described in Recursive’s OpenCode guide, and Pi can use Radius with /login radius. Pi labels Radius as “Flint (free, trains on data)”. Direct API access is limited to select partners.

Settings that matter

Recursive’s generation settings page recommends temperature 1.0, top-p 0.95 and top-k 20. Because top_k is not a standard OpenAI field, pass it through extra_body where your endpoint supports it.

Flint’s reasoning is always on. Do not send a reasoning-effort setting or a switch to disable thinking. The model card advises an output budget of at least 32,768 tokens for hard problems and keeping returned reasoning_content in the conversation history. For OpenCode, Recursive suggests a 65,536-token output budget where supported, automatic compaction with a matching buffer, session sharing off, and the same Flint model for main and auxiliary agents. Our OpenCode 2.2 guide explains where those settings live.

A good first task, Recursive suggests, is to ask Flint to explain one function in a small repository you know, add a focused test, and then fix what the test exposes.

Limitations

  • Text only. Flint does not read images or audio.
  • Long reasoning. Reasoning can add latency and use up the output budget before the answer finishes. If a reply stops with length, raise the budget.
  • Context quality. The blog says Flint supports up to 512k tokens but recommends staying within 256k. Recursive notes the context limit is a capacity figure, not a promise of equal accuracy at every length.
  • Transformers support. The model card says Transformers modeling code is not included yet in the main repository.
  • Safety. Recursive warns the model can produce faulty code or unsafe tool calls, so tool-using apps should enforce their own permissions.

Frequently asked questions

What is Recursive Flint?

Flint is Recursive’s open-weight mixture-of-experts model for coding agents. It has 39.8 billion total parameters, activates about 2.5 billion per token, handles text in and text out with reasoning and tool calls, and is released under the Apache 2.0 license.

What hardware do I need to run Flint locally?

Recursive’s rec engine needs an Apple Silicon Mac with an M4 Pro, M5 Pro or M5 Max, at least 48 GB of unified memory and about 30 GB of free disk space. Other machines can use Ollama, LM Studio or llama.cpp with a compatible quantized file.

What is Flint’s context window?

The model card documents a 262,144-token configuration that covers both prompt and output. Recursive’s blog says Flint supports up to 512k tokens but recommends staying within 256k tokens for the best results.

Is there a hosted Flint API?

Direct Flint API access is limited to select partners, according to Recursive’s FAQ. For free hosted use, Recursive points to OpenCode’s free tier (Flint-40B) and Radius inside the Pi harness, which Pi labels as training on your data.

What sampling settings should I use with Flint?

Recursive recommends temperature 1.0, top-p 0.95 and top-k 20. Reasoning is always on, so do not send a reasoning-effort or disable-thinking setting, and allow at least 32,768 output tokens for hard problems.

Sources

Image: Flint logo by Recursive, from the official Flint model repository on Hugging Face. FreeAnswers is not affiliated with Recursive.

Related answers

Leave a Comment