Recursive Flint: The Free Coding Model Built by AI (2026)

A $4.65 billion startup just let its own AI help build a coding model in under four months — and then gave the result away for free. Recursive's Flint is a 40-billion-parameter coding model designed to run entirely on your laptop, offline, with no subscription and no per-token fees. Here is what it actually is, what the verified numbers say, and how it honestly compares to the paid coding assistants you might already pay for.

What Flint actually is

Flint, released on October 21, 2026, is Recursive's first public model. It is a mixture-of-experts (MoE) coding model: instead of using all 40 billion parameters for every task, it activates only about 2.5 billion per task — roughly like keeping a huge toolbox but only pulling out the tools you need. That sparsity is what makes it small enough to run locally.

Key specs, based on the launch reporting and technical coverage of the model card:

• 40B total parameters, ~2.5B active per task — free to download on Hugging Face under the Apache 2.0 license, which permits free use and modification.
• Runs on 48GB Apple Silicon Macs or a 48GB NVIDIA GPU — the 4-bit weights occupy roughly 23GB. Wired's Will Knight tested it on a Mac with 48GB of memory.
• Context window: 262,144 tokens per the model card, up to 512k per Recursive's blog — treat the model card as the authority here, not the marketing.
• Agent-oriented: text-only, with reasoning support and OpenAI-compatible tool calls, aimed at coding agents that investigate codebases, fix bugs, and write tests.

The company behind it, Recursive, emerged from stealth in May 2026 with $650 million in funding at a $4.65 billion valuation, backed by GV, Nvidia, and AMD Ventures. It is led by CEO Richard Socher, Salesforce's former chief scientist, and its stated mission is recursive self-improvement: AI that builds better AI.

Built by AI: the Eureka Machine story

This is the part that makes Flint more than just another open model. Recursive says its automated research system, the "Eureka Machine," helped with nearly every step of Flint's development — from pretraining through post-training to inference — letting a small team build the model from scratch in under four months.

The company credits Eureka with three concrete contributions: a sparser model architecture, low-bit training that shrank Flint by 71% with "almost no loss in quality," and custom compute kernels that speed up the underlying math.

Be honest about what this proves. It is an early, limited test of recursive self-improvement — AI producing a laptop-scale model under human oversight — not AI redesigning itself unsupervised. Recursive isn't alone in leaning on AI for research: OpenAI said that by mid-August 2026, its AI agents were logging 3.1 workdays for every one worked by a human researcher. The direction is real; the headline deserves the caveat.

How fast is it, really? The numbers, labeled

All benchmark and speed figures below are company-reported unless stated otherwise. Independent benchmarks for Flint do not exist yet, and that matters.

• 131 tokens per second on a MacBook Pro with Apple's M4 Pro chip — roughly double the 58–69 tokens/sec Recursive measured for comparable models on the same machine, including Google's Gemma 4 and Nvidia's Nemotron 3.5 Lightning. On Nvidia's RTX Spark, it topped 150 tokens/sec.
• SWE-Bench Verified: Flint outperformed other models under 50 billion parameters on this real-world bug-fixing benchmark. On Recursive's own six-test average, only one model scored higher — Alibaba's Qwen3.8 27B.
• Technical coverage adds the detail that the picture is split: 95.1% on IFEval and 74.2% on SWE-bench Verified look strong, but on Terminal-Bench 2.1, Flint scored 54.9% against 73.0% for the dense Qwen3.8. Agentic coding is its strength; terminal-style tasks are the weak spot.

The one independent datapoint available is encouraging: Wired's Will Knight ran Flint on a 48GB Mac and built several small programs — a to-do list app and a web-research tool — finding that Flint "nailed his coding requests every time," where other small coding models he had tried hit errors or ran slowly compared with Claude or Codex. One reviewer's experience is not a benchmark, but it is a genuine signal.

Flint vs paid coding assistants: the honest cost comparison

This is the commercial question everyone actually asks: can a free local model replace a paid subscription? Here is the current price landscape, verified as of October 2026:

| Tool | Price | Model | |---|---|---| | Recursive Flint | $0, one-time download | Local 40B MoE, runs offline, Apache 2.0 | | GitHub Copilot | Free / $10/mo Pro / $39/mo Pro+ / $100/mo Max | Cloud, multi-model; completions unlimited on paid plans, agent usage draws on AI credits since June 2026 | | Cursor | Hobby free (limited) / $20/mo Pro / $60/mo Pro+ / $200/mo Ultra | Cloud agent + editor; Pro includes ~$20 of usage, heavier agent users land at $60–100+/mo | | Claude Code | From $20/mo (Claude Pro) | Cloud, Anthropic's models | | ChatGPT Codex | From $20/mo (ChatGPT Plus) | Cloud, bundled with the rest of ChatGPT |

Pricing details come from GitHub's own pricing announcement, Cursor's pricing page as checked by 2026 comparison roundups, and the same roundups for Claude and ChatGPT tiers. Prices move — always confirm at checkout.

The honest math: a developer on Cursor Pro pays $240/year; on Copilot Pro, $120/year. Flint's subscription cost is $0 forever — but it demands hardware you may not own (a 48GB Mac or a 48GB NVIDIA GPU costs real money), and Recursive itself admits Flint needs more guidance than the big paid assistants. Their advice: break large tasks into smaller steps. If you want one-click "just do it" agentic coding from a weaker laptop, the paid cloud tools still earn their fee. If you have the hardware and want privacy, zero marginal cost, and offline work, Flint changes the equation completely.

Who should try it — and who should skip it

Try Flint if: you code on a recent high-memory Mac or a strong NVIDIA GPU; you work with proprietary or client code you would rather not send to a cloud provider; you are a student or indie builder who wants a capable coding agent at zero subscription cost; or you want to experiment with open-weight models without renting cloud GPUs.

Skip it for now if: your machine has 16GB of memory (it will not fit); you want a polished one-click assistant and do not want to manage weights and local inference tooling; or your team needs centralized billing, SSO, and audit controls — the paid enterprise tiers exist precisely for that.

TechCrunch's own comparison set Flint against Meta's Muse Spark Open — the 816B-parameter open model from a week earlier that needs a terabyte of memory. Two open-weight releases at opposite ends of the spectrum, plus the cheap-AI-agents trend we covered with DeepSeek's V4.1-Flash: the industry is splitting into two races, bigger cloud brains and smaller local ones.

How to get started

Flint's weights are on Hugging Face under Apache 2.0 — free to download, use, and modify. You will need a 48GB Apple Silicon Mac or a 48GB NVIDIA GPU, ~23GB of disk space for the 4-bit weights, and local inference tooling that supports mixture-of-experts models. Expect setup friction: this is a research-grade release, not a consumer app. Check the current model card before relying on it.

One editorial note: this blog has not run Flint ourselves, so we are not claiming hands-on experience. Everything above is either company-reported (labeled as such), from the model card via technical coverage, or from Wired's independent hands-on — all linked below.

What to watch next

Recursive's longer-term goal, per CEO Socher, is full recursive self-improvement under human oversight — each model making the Eureka Machine a better researcher. The real tests: whether independent benchmarks confirm the speed and SWE-bench claims, whether developers build real tooling around Flint, and whether a $4.65B startup can keep giving away models while paying for compute (it already has a $410M AWS compute deal). The weights you download now are yours under Apache 2.0 regardless.

Sources: the launch story was broken by TechCrunch (Oct 21, 2026, updated Oct 22); model-card specs and benchmark detail via MarkTechPost; GitHub Copilot pricing from GitHub's official announcement; Cursor and rival pricing cross-checked via Spectrum AI Lab's 2026 roundup and Depizo's Copilot-vs-ChatGPT breakdown.

No affiliate links in this article — tool names are plain text. Benchmark and speed figures are company-reported unless noted.