Laguna S 2.1 — poolside's open-weight coding model
Laguna S 2.1
poolside just released a coding AI you can download and run yourself, for free. It's called Laguna S 2.1, and it holds its own against models several times its size on real programming work. The surprising part isn't how big it is. It's how it learned.
118 billion numbers, 8 billion awake per word
Laguna is a mixture-of-experts model. Instead of one big network, it's built as hundreds of small specialist sub-networks. Together they hold a hundred and eighteen billion internal numbers, the values it learned in training. But for each word, the model wakes only about eight billion of them, routing that word to a handful of specialists instead of running the whole thing. poolside calls it the West's most capable open model in its class. That's poolside's own claim.
It practiced on about 409,000 real software tasks
Here's what's new. The coding skill didn't come from size. It came from practice. poolside had the model attempt real software tasks, about four hundred and nine thousand of them. The biggest batch was thirty eight thousand real code changes, pulled from seventeen thousand open-source projects. The model tried each task, and poolside rewarded it only when the result actually worked. It's like an apprentice who learns by doing the job, not by reading about it.
Letting it think between actions more than doubles its bug-fixing
Learning by doing comes down to two things. The first is the training. The model wrote code, ran it, and kept the version that passed, the one that matched what a real developer had committed. The reward came from code that actually ran, not from text describing it. The second is what happens when you actually run the model, and it's the surprising part. Between each step, it thinks. It reasons privately, runs a command, reads the result, then reasons again. That private reasoning is where most of the skill turns out to live. Turn the thinking on versus off, on a test that has it fix real GitHub bugs, and the score more than doubles, from sixteen point five to forty point four. On finishing real terminal tasks, it climbs from about sixty to seventy. That thinking isn't free. It costs about two and a half times as many words, roughly two hundred and fifty thousand instead of a hundred thousand on the bug-fixing test. But the gain is real.
Tops multilingual bug-fixing at 8B active — frontier still ahead
Tops multilingual bug-fixing at 8B active, frontier still ahead
There's one test where Laguna comes out on top. It's called SWE-bench Multilingual. It has models fix real GitHub issues across many programming languages. Laguna scores seventy eight point five, the best in the group. It edges out Qwen 3.7 Max and beats every larger open model: DeepSeek at forty nine billion active parameters, Tencent's Hy3 at twenty one billion, Nvidia's Nemotron at fifty five billion. Laguna does it on eight billion. So it's best in its weight class at this kind of coding, not best overall. Elsewhere it sits mid-pack, and the closed frontier, models like Claude Fable 5 and Kimi K3, is clearly ahead. These are poolside's own scores, each rival measured at its best public number.
Open, ungated, and free to build on
Here's the part that matters if you build. Laguna ships under a license called open M D W: anyone can use it, change it, even sell products built on it, for free, and there's no gate to accept. poolside says a compressed, four-bit version fits on a single desktop-class machine. That claim hasn't been verified outside poolside yet. If it holds, your code never has to leave the building to get this kind of help.
The lever wasn't a bigger model
The lever wasn't a bigger model. It was practice, and room to think.
What made Laguna good at coding wasn't a bigger model. It was practice on real tasks, and room to think between each step. Two things are worth watching: whether the closed frontier keeps its lead once other labs train this way, and whether an outside leaderboard reproduces that multilingual win. Every score today comes from poolside's own testing, and an independent run of the weights hasn't happened yet.
