October 28, 2026

Recursive Flint: A Laptop-Sized Coding Model, Built (Mostly) by AI

Recursive dropped Flint this week: a free, Apache 2.0, 40B MoE coding model that runs on a laptop and was largely built by its own AI research system. I dug int...

Recursive released Flint on October 21: a free, open-weight coding model that runs on a laptop. The headline isn’t the size, the license, or even the benchmarks — it’s the origin story. Recursive says Flint was largely built by AI, using its automated research system, the Eureka Machine, across nearly every stage of the pipeline. A small human team took it from scratch to release in under four months.

This is the first commercial release built through recursive self-improvement, and it’s the kind of thing that changes what “open model” means. Here’s the engineering read.

What Flint actually is

Flint is a decoder-only mixture-of-experts model:

  • 39.8B total parameters, ~2.5B active per token (about 2.2B active body plus output head)
  • 49 layers: 2 dense, then 47 MoE layers with 256 routed experts each, 6 selected per token plus one always-on shared expert
  • 262,144-token context window (Recursive recommends budgeting at least 32K output tokens for hard problems)
  • Apache 2.0, weights on Hugging Face as RecursiveSI/flint; BF16 checkpoint is ~80 GB, with quantized builds available
  • Runs offline, keeps your code private, and costs nothing per token

The MoE shape is the whole trick: because only ~2.5B params are active per token, each token reads only ~1.4 GB of weights at 4.5 bits. On an M4 Pro’s 273 GB/s memory bandwidth, that puts the theoretical decode ceiling around 194 tokens/s — and Recursive reports ~131 tokens/s in practice, roughly double what Gemma 4 and Nemotron 3.5 Lightning managed on the same machine.

Running it

You don’t need Recursive’s custom rec engine to use it — standard tooling works:

# Ollama (upgrade to 0.40.9+ first)
ollama pull flint
ollama run flint

# Or point Claude Code at it via Ollama
ollama launch claude --model flint

llama.cpp, LM Studio, vLLM, and SGLang are all supported. Recursive ships a single B300/H200 serving kit for self-hosting, but the whole point is that a 16GB+ Mac or a 24GB GPU is enough. That’s a sharp contrast with Meta’s Muse Spark Open, released the same week at 816B parameters and a terabyte of memory.

The benchmarks, with a grain of salt

Recursive’s published numbers (vendor-reported, so treat them as directional):

  • 89.7% on SWE-bench Pro v2
  • 74.2% on SWE-bench Verified — best among sub-50B models, per TechCrunch
  • 54.9% on Terminal-Bench 2.1, where Recursive counts any solution reached via web retrieval as a failure — a stricter standard than most

On Recursive’s own six-test average, only Alibaba’s Qwen3.8 27B scored higher. Wired’s Will Knight ran it on a 48GB Mac and found it nailed small programs (a to-do app, a research scraper) where similar local models stalled or errored.

What’s actually new here

Two things worth paying attention to:

  1. The size/quality tradeoff is collapsing. A 40B MoE that beats other sub-50B models and runs at interactive speed on a laptop is now table stakes for local agentic coding. Qwen3-Coder 30B is the incumbent to beat; Flint is a credible new contender, and it’s Apache 2.0 with unrestricted commercial use.

  2. The “built by AI” claim is the real story. The Eureka Machine contributed to data selection, optimizers, training recipes, infrastructure, and inference optimization — including a sparser model shape and low-bit training that shrank Flint by 71%. That’s AI-assisted model development escaping the “write a paper about it” phase into shipping. Company-reported, unverified, and still remarkable for a 2026 release.

What sucks

  • It needs more hand-holding. Recursive is upfront that Flint requires more explicit guidance than big paid assistants. Break large tasks into small subtasks, or it drifts. The reviewers’ wins were small, well-scoped programs — don’t expect repo-scale refactors on day one.
  • The speed headline is engine-dependent. 131 tok/s comes from Recursive’s custom engine on specific hardware; your llama.cpp quant may be slower. Measure on your own machine.
  • Benchmarks are publisher-reported. Compare Flint and Qwen3.8 on your repository before switching.
  • Commercial licensing is a question mark. Free for individuals today; enterprise licensing arrives in early 2027, pricing TBD.

Verdict

I’d use it — for the right jobs. Flint is a genuinely useful local fallback: offline, private, free, and fast enough for agentic workflows. It’s not replacing Claude Code for multi-file refactors, but as a privacy-preserving layer for day-to-day agentic coding, it’s the strongest new option on a laptop in months.

The deeper take: when a model’s trainer is itself a model, the “open weights” conversation stops being about licensing and starts being about provenance. That’s worth watching closely — and it shipped on a Tuesday.

Source: TechCrunch — “Recursive launches Flint, a laptop coding model it says was largely built by AI” (Oct 21, 2026).