Laguna M.1

What Laguna M.1 is

Laguna M.1 is a large language model from Poolside built for software engineering work. It has 225 billion total parameters, but only 23 billion of them are active for each token it generates^[1]. The gap between those two numbers is the signature of a mixture-of-experts (MoE) architecture: instead of passing every token through the whole network, the model routes each token through a small subset of specialised sub-networks, so the compute spent on a single step is closer to that of a much smaller dense model. Poolside makes this comparison explicitly, noting that the dense models it measures against activate more parameters per token^[1]. The trade-off is memory: every parameter still has to be stored and loaded, even though only a fraction is used at each step.

The target workload is what Poolside calls long-horizon agentic coding, rather than single-turn code completion^[1]. The distinction matters for how a model of this kind is used. A completion model answers one prompt with one block of code and stops. An agentic model is expected to work through a task over many steps — editing files, running commands, reading the results, correcting itself — so the demands fall as much on sustained coherence and tool use across a long session as on any single answer.

How it was trained

Poolside trained the model from scratch on 30 trillion tokens, using 6,144 interconnected Nvidia Hopper GPUs^[1]. Training from scratch means the weights were not initialised from an existing model; the network learned its parameters from the training data alone. Pre-training finished at the end of 2025^[1].

Benchmark results

On SWE-Bench Pro the model scores 46.9 percent, and on Terminal-Bench 2.0 it scores 40.7 percent^[1]. Both belong to the family of benchmarks that hand a model realistic software tasks and then check whether the resulting work is correct, rather than testing whether it can reproduce a short snippet. On SWE-bench Verified, the version of the benchmark most widely reported, Laguna M.1 resolves 72.5 percent of tasks; DeepSeek-V4-Flash resolves 79.0 percent and Claude Sonnet 4.6 resolves 79.6 percent^[1]. On that benchmark the model therefore trails both of the comparison systems.

Scores on different benchmarks are not directly comparable to one another, since each one builds its tasks, its environment and its pass criteria differently. What the numbers show is that a model can be strong enough to sit on the same table as the larger closed systems while still giving up several points to them on the most-quoted measure.

Release and open weights

When the model was announced, the weights were not open. Poolside listed open weights as coming and said that access could be requested in the meantime^[1]. The weights reached Hugging Face on 15 June 2026, seven weeks after the announcement^[2]. Open weights change who can use a model in practice: once the files are downloadable, outside teams can run the model on their own hardware, inspect it and fine-tune it, instead of reaching it only through a hosted interface. The seven-week gap also shaped the early period, during which the model could be tried by those Poolside chose to let in.

Where it sits in the Laguna family

Poolside describes Laguna M.1 as the foundation for the rest of the Laguna family^[1]. Positioned that way, it is less a finished product than the base layer of a line: the large MoE model supplies the capability, and the rest of the family is built on top of it. For readers following the field, that framing is the useful part — it says where the company intends the work to go, and it explains why so much of the training budget went into a single general-purpose coding model rather than a narrow one.