Skip to content
ML2
Coding & Shipping AgentsTool#Local / BYO model#Open source

Recursive Flint: a fast local coding model for agents

Added NewCopyable prompt below

Published Last verified Released

Our verdict

The most practical way right now to run a coding agent fully offline on a Mac with enough memory. Because only about 2.5B of its 40B parameters are active per token, builders are reporting 115 to 132 tokens a second on M5 Pro and M5 Max Macs, fast enough for everyday agent loops, and you can point Claude Code, OpenCode or Hermes Agent at it through Ollama. The benchmark scores are Recursive's own, and Recursive itself says it needs more guidance than the big closed models, so use it for scoped, private or high-volume tasks and keep a frontier model for the hard ones.

Who it's for
Builders with an Apple Silicon Mac (32 GB+, ideally 48 GB+) or a 12-24 GB NVIDIA GPU who want private, offline or zero-API-cost coding agents
Leverage
A local coding agent that is fast enough for real work, with your code never leaving your machine and no per-token fees
Effort
Low
Date added

What it does

Flint (RecursiveSI/flint on Hugging Face, Apache 2.0) is a decoder-only mixture-of-experts model built for coding agents: 39.8B total parameters, about 2.5B active per token, a documented 262,144-token context, reasoning and OpenAI-compatible tool calls. Recursive says its automated research system, the Eureka Machine, contributed to most of the stack and that the model was built from scratch in under four months. It ships as a BF16 checkpoint (about 80 GB, served with SGLang on one H200 or B300) plus quantized builds: an MLX INT4 build of about 22.6 GB for Apple Silicon and a 23 GB Ollama build with ollama launch recipes for Claude Code, OpenCode, Hermes Agent and OpenClaw. Recursive also released rec, an Apache-2.0 inference engine for M4 Pro, M5 Pro and M5 Max Macs with 48 GB+ that launches the pi or OpenCode harness and serves an OpenAI-compatible API on 127.0.0.1:11500. Recursive reports 54.9% on Terminal-Bench 2.1 and 74.2% on SWE-bench Verified; these are vendor-run numbers.

How builders use it

  • On a Mac, install Ollama and run ollama launch claude --model flint (or opencode / hermes) to drive your usual agent with the local model.
  • On an M4 Pro, M5 Pro or M5 Max with 48 GB+, install rec and run rec launch pi or rec launch opencode in the project folder; rec serve gives any tool an OpenAI-compatible endpoint.
  • Give it scoped tasks (one bug, one feature, one test file) with explicit instructions, as Recursive recommends, and stay within 256K tokens of context.
  • Keep a hosted frontier model for large refactors and compare results on your own repo before switching defaults.

What to expect

A coding agent that runs offline at interactive speed on your own hardware, with no API fees and no code sent to a vendor.

Copy this: Run Flint locally (Ollama or rec)

# Official sources: huggingface.co/RecursiveSI/flint (Apache 2.0), ollama.com/library/flint, github.com/recursive-oss/rec

# Option A: Ollama (about 23 GB download)
ollama run flint                        # quick chat test
ollama launch claude --model flint      # or: opencode / hermes / openclaw

# Option B: rec (Apple M4 Pro / M5 Pro / M5 Max, 48 GB+ unified memory, ~30 GB disk)
curl -fsSL https://raw.githubusercontent.com/recursive-oss/rec/main/install.sh | bash
rec doctor                              # checks your hardware first
rec launch pi                           # or: rec launch opencode
rec serve                               # OpenAI-compatible API at http://127.0.0.1:11500/v1, model "flint"

# Recommended sampling: temperature 1.0, top-p 0.95, top-k 20; keep context under 256K.

Related discoveries

See the newest picks →
  • ToolCoding & Shipping Agents

    Aider

    Open-source A.I. pair programming in your terminal, with git commits built in.

    Effort: Low#Local / BYO model#Open source

  • ToolCoding & Shipping Agents

    Cline

    Open-source autonomous coding agent for your IDE, CLI or as an SDK.

    Effort: Low#Plugin / extension#Local / BYO model

  • ToolCoding & Shipping Agents

    opencode

    An open-source coding agent for the terminal that works with many model providers.

    Effort: Low#Local / BYO model#Open source

  • ToolVibe-Coding & Prompt-to-App

    Continue

    Open-source coding agent for VS Code and JetBrains, any model.

    Effort: Medium#Plugin / extension#Local / BYO model

Want more like this?

Practical discoveries, without the noise.

Get the newsletterMore in Coding & Shipping Agents