The American AI company Poolside has released its in-house large language model "Laguna M.1" as an open model. The weights of its most capable model, which until now was only available through a paid API, can now be downloaded by anyone under the permissive Apache 2.0 license. A community-quantized build even runs locally on a Mac with 128GB of memory. In terms of performance, Laguna M.1 edges out a French model but still trails the fast-rising open models from China.

From Paid API to a Freely Available Top Model

Poolside is an American startup focused on AI for software development. On June 18, 2026, the company published the model weights for Laguna M.1, which it describes as its most capable model to date. The model handles a context length of up to 256K tokens, and both the pretrained base checkpoint and the post-trained checkpoint are distributed on the model-sharing site Hugging Face under the Apache 2.0 license.

Laguna M.1 is not a newcomer. Poolside first announced it on April 28, 2026, alongside the smaller Laguna XS.2. At launch, only the compact XS.2 was released as an open model, while the higher-end Laguna M.1 was offered as a paid product accessible via API. This release reverses that approach, opening up the internals of a top-tier model that had been a source of revenue. With many cutting-edge models reachable only through the cloud, it adds to the options for running advanced models on your own hardware.

A Mixture-of-Experts Design with 225B Total and 23B Active Parameters

Laguna M.1 is a large model with roughly 225 billion total parameters. However, only about 23 billion parameters are actually active during any single inference pass. This is because it uses a Mixture-of-Experts (MoE) architecture, which activates only the experts needed for a given input, aiming to combine the benefits of large scale with lighter processing.

The smaller Laguna XS.2 released at the same time has about 33 billion total and 3 billion active parameters, making it easier to run locally. The newly opened Laguna M.1 sits above it as the core model carrying the heavier performance load.

Beating a French Model but Falling Short of China's Best

Poolside also shared benchmark results comparing Laguna M.1 with other major models. The comparison set included the French Devstral 2 (a 123B dense model), the Chinese GLM-4.7, DeepSeek-V4-Flash, and Qwen3.5, as well as Anthropic's Claude Sonnet 4.6.

The results show Laguna M.1 outperforming the French Devstral 2. At the same time, it fell short on scores against the Chinese DeepSeek-V4-Flash and Qwen3.5, which have been gaining ground in the open-model space. The picture reflects how Western developers are chasing their Chinese counterparts in open models. It is worth keeping in mind, however, that many of these figures were published by the developers themselves or by various rankings.

A Quantized Build That Runs on a 128GB Mac

The community moved quickly after the release, and several quantized models have already appeared on Hugging Face. Among them, the 3-bit "Laguna-M.1-MLX-Q3" was reported to run locally on Apple Silicon Macs. Specifically, a working example showed it generating about 26 tokens per second with roughly 100GB of peak memory on a Mac equipped with an M3 Max and 128GB of unified memory.

It is an example of how even a giant model can run on a single high-performance Mac, rather than in a data center, by combining quantization with memory-saving techniques. That said, a Mac with 128GB of memory is an expensive configuration, so this is not something everyone can casually try. Even so, the fact that there is now more room to run a near-cutting-edge model on your own machine, without relying on external servers or API fees, is an appealing change for developers.

Summary

Poolside has released Laguna M.1, the top model it previously offered through a paid API, as an open model under the permissive Apache 2.0 license, keeping its full 225 billion parameters. Its performance surpasses a French model but realistically falls short of the surging open models from China. On the other hand, the fact that a quantized build runs locally on a 128GB Mac is hard to overlook for developers who want to use a high-performance model without depending on the cloud. As the West and China compete for leadership in open models, the addition of one more option is no small matter.