Robert McHardy @ ICLR 2026 🏖️

Super excited to release the weights of Laguna XS.2 under Apache 2.0 to the community today. We are also opening access to our most capable model, Laguna M.1, via our API. Both models are designed for long-horizon agentic coding with strong performance on SWE-bench-style tasks. Learn more on our blog.

@teortaxesTex XS.2 was 20 days flat in pre-training, not pure training time though since we paused for a bit to run some additional experiments. Iirc it was 14 days pure training time (and it was 30T tokens, not 33T)

225B total 23B activated. 6-8K H200s 30B total 3B active, 33T tokens: 20-something days on 2K Man, this really puts the V3-671B's feat (55 days) into perspective a year and a half ago, those guys could push H800s to the level H200s normally can't touch (they're improving though)

Here is a 2nd batch of April architecture drops. What a month! - Ant Ling 2.6 1T - Minimax M2.7 - Xiaomi MiMo V2.5 - Poolside Laguna XS.2 - Tencent Hy3-preview - IBM Granite 4.1

We are an American company with a global team and global aspirations. The story of @poolsideai is that early on in the life of the company we decided to focus on building out our applied research org. in Europe. That’s been the seed for an amazing team and a competitive advantage. Today we have team members all over the world, Europe and US are roughly equal in size, Asia is growing. Three years ago we thought France would be a great place to build from but in the early days found our hiring happened all across Europe instead. Today we have less than a handful of folks in France but large teams in Europe in London, Amsterdam, Zurich etc. We operate as a remote first company but have an office in Paris where we do monthly on-sites (it is great logistics for this) and an office in London which is used on a daily basis. When we raised capital early on, the vast majority of it came from US investors but European (including French ones) have been a part of our rounds. When we got to France (almost 3 years ago) we were offered significant double digit million research grants. My cofounder @jasoncwarner and I did not feel comfortable accepting the grants when we realized that France was going to be only a small part of our story. So we respectfully turned them down. France keeps a special place in our hearts but we’re a global company with global aspirations.

@eliebakouch We actually didn't have any problems with the routers and this argument is/was unused in training, but nice hypothesis!

nice oss release from poolside! i really like the arch :) some "unusual" hidden modeling details: > they use per head gated attention like qwen or stepfun, people usually use sigmoid, but they use softplus! they also use it on both swa and full attention which is nice! this is also why they don't use sink > there is a "moe_router_logit_softcapping" parameter in the config set to 0 by default, but this is non-standard i think, means that they likely saw some instabilities in the router scores? this also changes the bias strength in the top-k since they do it after > more attention heads on swa layers, xiaomi v2 flash did higher kv heads for swa layers > they use a different beta_fast for yarn than most models (64 instead of 32) and partial (half) rope for the full attention layers

@erikqu_ We mean starting from random weights by from scratch if that answers your question

I wish there existed a method to compare what “training from scratch” truly means

@andrew_n_carr @sakurayukiai Our implementation for this is actually from June last year or so, back then no correct & efficient open-source implementations existed. I remember one that was very fast but only so due to a total of 4 (iirc) different race conditions in the code

@sakurayukiai Yeah, this is legit stuff

Wait, Poolside trained a 225B MoE from scratch using the Muon optimizer instead of AdamW?? They distributed the Newton-Schulz math across GPU ranks to keep overhead under 1%. 15% faster convergence and half the optimizer VRAM. I need to see this training code immediately 👀

Some more amazing models launched today! Congrats to @poolsideai on the release of Laguna XS.2 and Laguna M.1

First open release by Poolside (upcoming LLM lab) 🙌 Coming in as #12 on SWE-Bench Pro, not too far behind Qwen3.6 🔥

In my short time at Poolside I was involved in 3 model launches (one internal), and it is impressive how industrialized the process is The XS release is most iconic: from nothing to release in 5 weeks

The first public foundation models from @poolsideai just dropped on OpenRouter! Laguna M.1 and Laguna XS.2. Built from scratch for agentic coding and long-horizon work. Free for a limited time ⬇️

@rasbt April isn't over yet, @rasbt 😄 x.com/poolsideai/sta…

April was a pretty strong month for LLM releases: - Gemma 4 - GLM-5.1 - Qwen3.6 - Kimi K2.6 - DeepSeek V4 All are now added to the LLM Architecture Gallery. More details once I am fully back in May!

@PMinervini You should check out Laguna XS.2, Pasquale! Fresh out the oven x.com/poolsideai/sta…

Some quick stats on what local models may work for you if you are an NLP/ML person:

@dejavucoder @poolsideai that's the plan indeed!

@poolsideai looks like a very good release. do you plan to publish a tech report?

Today we’re releasing Laguna XS.2, Poolside’s first open-weight model. It’s a 33B total / 3B active MoE model built for agentic coding and long-horizon tasks. Trained fully in-house on our own stack. Runs on a single GPU. Released under Apache 2.0. Links 👇 Weights: huggingface.co/poolside/Lagun… API: platform.poolside.ai Blog: poolside.ai/blog/laguna-a-…

@ctnzr Thank you, Bryan! 💪 Very excited to share this model with the world

Love to see more open models coming out! Great work from Poolside

Today we’re releasing Laguna M.1 and Laguna XS.2, our first public models. Laguna XS.2 is our first open-weight release, with weights available today on Hugging Face: huggingface.co/poolside/Lagun… A few details on what went into them: large-scale pre-training, data mixture optimization, synthetic data, optimizer efficiency, and async agent RL.

@poolsideai Let's go! 🔥

Today we’re shipping Laguna M.1 and Laguna XS.2 – our first public models. We’re also shipping our agent harness and a preview product experience. Both models were trained from scratch on our own stack: data pipelines, training infrastructure, and agent RL.

Poolside is heading to ICLR 2026 (@iclr_conf) in Rio this week 🇧🇷 @robert_mchardy , @iamgrigorev, @marah_i_abdin, @helloo_polly, @maxwegman, @SzymonOzog_ and @ArkadiiBessonov will be there from the team. If you’re working on training infra, synthetic data, coding models, RL, or anything at the frontier — come find us!

Introducing Target Policy Optimization (TPO): TPO turns GRPO into supervised learning: build a target distribution over sampled completions, then fit with cross-entropy. The gradient vanishes once the target is matched, making multi-epoch training smooth. 🧵(1/4)