Skip to content
AI STACK SIGNAL

Signals/ AI Agents/ Brief /

Flint is a free coding model small enough for a laptop

Recursive released a 40B open model that activates only 2.5B parameters at a time, with benchmark scores it reports itself.

Recursive, an AI lab, released Flint on October 21. It is its first model: an open-weight coding model under the Apache 2.0 license.

Flint has about 39.8 billion parameters in total but uses only about 2.5 billion for each token it produces. That design is called mixture of experts, and it keeps the model light enough to run locally. AI Weekly reports quantized versions down to 11.4 GB and roughly 125 tokens per second on an M4 Pro MacBook. The context window is 262,000 tokens.

Recursive reports 89.7% on SWE-bench Pro v2, 74.2% on SWE-bench Verified and 54.9% on Terminal-Bench 2.1. It says most of the model’s development was automated by its own “Eureka Machine” research system.

The catch: Every score and speed figure comes from Recursive, and neither source tested it independently. RuntimeWire does not list a download location, while AI Weekly points to a Hugging Face page.

Should you switch?

  • Maybe yes: you write code, care about privacy, and want a free model to compare against your current tool.
  • Probably not yet: you need proven results on your own codebase, or you do not have a Mac or PC with enough memory.

Prompts to test it

Read the function below. Find the bug, explain it in two sentences, then give the fixed version and one test that would have caught it.

[paste code]

Sources & links 2

  1. AI Weekly: Recursive opens Flint, a 40B/2.5B MoE coder aiweekly.co ↗
  2. RuntimeWire: Recursive ships Flint, a local coding model runtimewire.com ↗

More on AI Agent News

More signals

Reel

A Website Is Probably Showing Strangers Your Home Address. This AI Skill Files the Opt-Outs

Unbroker is an open source skill for Hermes Agent. It finds your listings on people-search sites like Spokeo, Whitepages and BeenVerified, checks they're really you, files opt-out forms and emails, and re-checks later. Here's how it works, the install command, what it asks for, what still needs you, the privacy trade-offs, and a do-it-yourself checklist.

Helps People in the US who found their home address, phone number or relatives on a people-search site and want the removal paperwork handled, plus anyone who'd rather do it by hand with a clear checklist

  • Privacy
  • Data Brokers
  • Hermes Agent
Reel

Your Most Expensive AI Shouldn't Be the One Typing the Code

One developer made GPT-6.1 Sol the boss and a self-hosted GLM-5.3-Flash the worker inside OpenCode. The boss plans, briefs, reads every change and runs the checks; the worker explores, edits, tests and fixes. Here's the setup in plain English, the four agents, a simplified config, his honest warnings, and our tips for trying a smaller version.

Helps Developers and small teams who pay for an AI coding plan or API and want the expensive model to spend its effort on judgment, plus anyone with a spare GPU curious whether a local model can do the routine coding work

  • AI Coding
  • OpenCode
  • Local AI
Reel

50 AIs Were Told to Build a Thriving Civilization. They Built a Bureaucracy

YC Labs and Olam Labs put 50 AI agents in a survival Minecraft world for three hours with one goal: build a thriving civilization. They got a leader picked by list order, a food ration ledger, permits, and a lot of hiding at night. Here's what happened, why, what it means if you use AI helpers at work, and three prompts to run a team of them better.

Helps Anyone who uses AI helpers or agents at work, or is thinking about it, and wants to know how a team of AIs behaves when the goal is vague

  • AI Agents
  • AI Research
  • Open Source