Breaking

Sakana AI's Tiny Router TRINITY Tops GPT-5 on Code Generation

July 3, 2026 at 07:46 EDT

Foundation Models
AI Agents
Research & Papers

Sakana AI has unveiled TRINITY, in which a "tiny coordinator" built from a roughly 0.6B-parameter model and a lightweight head of about 10,000 parameters orchestrates multiple LLMs to outperform a single frontier model. The paper was first submitted to arXiv in December 2025, with an updated version released in April 2026, and is slated for acceptance at ICLR 2026. Details were disclosed on Sakana AI's official page.

Dec 2025 · Sakana AI · ICLR 2026

A 0.6B "coordinator" conducts frontier LLMs — and beats every one of them

TRINITY pairs a tiny language model with a ~10K-parameter head to dynamically route a pool of top models, assigning each a role turn by turn. The result outperforms GPT-5, Gemini 2.5 Pro and Claude-4-Sonnet used alone.

86.2%

pass@1 on LiveCodeBench — state-of-the-art at release

-21.9%

relative error vs 2nd-best, across 4 benchmarks

<20K

trainable params in the coordinator head

The manager is tiny — the models it directs are huge

Only the coordinator is trained. It steers frontier LLMs at inference — no weight merging, no retraining of the pool.

0.6B

Coordinator SLM

100B+

Frontier models in pool

One query, three roles, looped until verified

Each turn the coordinator picks one LLM from the pool and hands it a role.

Thinker

strategy & decomposition

→

Worker

concrete problem solving

→

Verifier

checks & completeness

↺ loops until the Verifier passes the answer or the turn limit is reached

How it was trained

RL and imitation learning failed under low signal-to-noise. TRINITY instead uses sep-CMA-ES, a derivative-free evolutionary strategy — effective on a tiny budget.

1.5k – 40k evaluations

total training budget

Why it excites developers

Learned orchestration reads context, not static rules
Balances cost & quality — not every task hits a frontier model
Beat all pool members zero-shot on unseen OOD tasks

Open questions

How different from existing routing platforms?
Hand-built multi-agent setups can still win in cases
Pool diversity, context length & recursive-scaling cost

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.