Laguna M.1

Poolside's 225B coding model with 23B active and a 256K context. Apache 2.0, open weights since June 2026, with a base checkpoint.

In short. Laguna M.1 is the largest open model from Poolside, built for coding agents. It has 225 billion parameters and uses 23 billion for each token. Poolside kept it behind its own API from April 2026 and released the weights, plus a base version for researchers, on 18 June 2026 under Apache 2.0.

Summary

Poolside introduced Laguna M.1 through its API on 28 April 2026 and published the weights on 18 June 2026 under Apache 2.0, both the post-trained model and a pre-trained base checkpoint. Poolside says it was pre-trained in 2025. The model is a 225.8B-parameter mixture of experts with 23.4B active parameters: 70 layers, the first three dense and 67 sparse with 256 experts, 16 of them routed per token plus one shared expert, and global attention in every layer with 64 query heads and 8 key-value heads. It uses RoPE with YaRN, a 262,144-token context and interleaved thinking between tool calls that you can switch off per request. With the weights release Poolside said that open weights are now its default.

In Poolside's own table, Laguna M.1 scores 74.6% on SWE-bench Verified, 63.1% on SWE-bench Multilingual, 49.2% on SWE-Bench Pro and 45.8% on Terminal-Bench 2.0. In the same table DeepSeek-V4 Flash (284B) scores 79.0%, 73.3%, 52.6% and 56.9%, and Qwen3.5-397B-A17B scores 76.2%, 69.3%, 50.9% and 52.5%, so M.1 trails both. It is ahead of Devstral 2 (123B dense) on the three scores that table reports, and ahead of GLM-4.7 on SWE-bench Verified and Terminal-Bench 2.0 while behind it on SWE-bench Multilingual (66.7%). Poolside's newer Laguna S 2.1, at about half the size, reports 78.5% on SWE-bench Multilingual and 59.4% on SWE-Bench Pro in its own table, well above M.1's scores on the same two benchmarks. All scores are Poolside's own, from its agent harness.

The BF16 weights are 451.6 GB, with FP8 at 238.6 GB and NVFP4 at 139.3 GB. The card lists support in vLLM, SGLang, TRT-LLM and Transformers 5.7.0 or later, and recommends vLLM or SGLang for single-node serving because loading the BF16 checkpoint in Transformers takes substantial multi-GPU memory. In the 30 days to 11 November 2026, Hugging Face counted about 774,000 downloads of the main repository, though download counts can include automated pulls.

Good for

  • Self-hosted coding agents on a multi-GPU server, under a plain Apache 2.0 license
  • Researchers who want both the base and the post-trained checkpoints of a 225B model
  • Long agent runs that fit in a 256K context
  • Commercial products, since Apache 2.0 allows them with only light conditions

Watch out for

  • It needs a multi-GPU node, because even the NVFP4 files are 139.3 GB
  • In Poolside's own table DeepSeek-V4 Flash and Qwen3.5-397B-A17B score higher on all four benchmarks
  • Poolside's newer S 2.1 reports higher scores at about half the size, so M.1 mainly stands out for its Apache 2.0 license and its larger size
  • It was pre-trained in 2025, so its data and recipe are older than the rest of the Laguna line

Performance claims come from the developer and have not been independently verified.

How much memory does it need?

Estimates for the Laguna-M.1-NVFP4 variant. A mixture-of-experts model must hold all of its experts in memory even though only some are used for each token, so total parameters decide the memory and active parameters decide the speed.

Variants Parameters BF16 FP8 / INT8 4-bit (Q4) Fits on
Laguna-M.1 (BF16) BF16 226B 23.4B active 519 GB 260 GB 156 GB Big-memory desktop
Laguna-M.1-FP8 FP8 226B 23.4B active 519 GB 260 GB 156 GB Big-memory desktop
Laguna-M.1-NVFP4 NVFP4 226B 23.4B active 519 GB 260 GB 153 GB Big-memory desktop

Estimates: weights plus about 15% overhead. Real usage grows with context length and depends on the software.

VRAM calculator

The license in plain words

One of the most widely used permissive licenses. You may use, change and sell products built on the model with only light conditions.

More about Apache 2.0

See also

Frequently asked questions

What is Laguna M.1?
Laguna M.1 is the largest open model from Poolside, built for coding agents. It has 225 billion parameters and uses 23 billion for each token. Poolside kept it behind its own API from April 2026 and released the weights, plus a base version for researchers, on 18 June 2026 under Apache 2.0.
Can I use Laguna M.1 commercially?
Yes. Laguna M.1 is released under Apache 2.0, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
Is Laguna M.1 open source?
Laguna M.1 is released under Apache 2.0, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
Can I run Laguna M.1 on my own computer?
Only on a machine with a lot of memory. The smallest variant, Laguna-M.1-NVFP4, needs about 153 GB at 4-bit precision. Desktops with 128 to 512 GB of unified memory can hold it, at modest speed.
When was Laguna M.1 released?
Laguna M.1 was released by Poolside on June 18, 2026.
How long a context window does Laguna M.1 have?
Laguna M.1 supports a context window of up to 256K tokens.

Official links

Sources

Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.

Release date and facts last verified November 11, 2026.

Related models

Model Released Parameters Context License Smallest variant needs
Laguna S 2.1 Poolside 118B 8B active 1M OpenMDW-1.1 (Permissive) 81 GB Workstation
Laguna XS 2.1 Poolside 33.4B 3B active 256K OpenMDW-1.1 (Permissive) 22 GB Gaming GPU
DeepSeek V4.1 Flash DeepSeek 552B 16B active 1M MIT (Permissive) 561 GB GPU server
Qwen3.5 Alibaba Qwen 873M – 403B 17B active 256K Apache 2.0 (Permissive) 0.602 GB Laptop or phone