RecursionAI

Research

Bootstrapping the frontier, our monthly R&D publication, and case studies from real deployments.

No theory for theory's sake. Everything here comes from real tests, real deployments, and products we're building.

Bootstrapped Frontier

Bootstrapping on the AI Frontier

In one of the most aggressive, highly funded industries, we're building a frontier systems lab our own way without the VC money.

Builders

Jackson & Ryker built the stack themselves

We developed Courier OS, built Courier Cloud from scratch, and deployed Scout — the local agent stack that runs on your Mac. No outsourced substrate. The same systems we research are the systems we ship.

  • Courier OS

    Local-first production runtime on Apple Silicon

  • Courier Cloud

    Same software, rented as flat-rate APIs

  • Scout

    Free chat, Pro code & automations — deployed

Courier vs Ollama

Head-to-head tool-calling — measured

Same M5 MacBook Pro. Same Gemma 4 26B A4B (matched 4-bit). Official BFCL scoring. Courier wins on speed and tool-calling accuracy — including when Ollama is allowed to think.

Hardware
M5 MacBook Pro
Measured
2026-07-22

Faster

+37% decode · 2.3× sooner first token

123.8 vs 90.1 tok/s decode. Time to first token 0.122s vs 0.277s — on the same 26B 4-bit weights.

Courier

123.8 tok/s

Ollama

90.1 tok/s

More accurate

Better tool-calling — even when Ollama thinks

Courier wins every matched mode on the 26B. With thinking on, Courier clears 90% (0.902) on BFCL's AST tool-calling categories — still ahead of Ollama-thinking at 0.874.

ModeCourierOllama
No-think0.8970.850
Thinking0.902>90%0.874

Structure holds

+13.5 pts on parallel tool calls

Parallel category 0.830 vs 0.695 matched no-think — where structured tool-calling reliability matters most.

Courier

0.830

Ollama (no-think)

0.695

Read the full study

Scorer: BFCL v4 official ast_checker. Function-calling path only — native OpenAI tool_calls scored by Berkeley's checker.