Prime Intellect Offers Free RL Fine-Tuning Runs on Poolside's Laguna XS.2
Prime Intellect is giving away free RL fine-tuning runs on Poolside's Laguna XS.2, a 33B MoE coding model that scores 68.2% on SWE-bench.
Takeaways
- Free RL training on Laguna XS.2: Prime Intellect is offering free Hosted Training runs on Poolside's 33B MoE coding model, first-come first-served.
- Strong benchmark numbers: Laguna XS.2 scores 68.2% on SWE-bench Verified and 44.5% on SWE-bench Pro, competitive with much larger models.
- Efficient MoE architecture: Only 3B of 33B parameters activate per token; runs on a single GPU or a Mac with 36 GB RAM via Ollama.
- Apache 2.0, fully open weights: Available on Hugging Face with no commercial restrictions.
- No GPU management needed: Prime Intellect Lab handles orchestration, rollouts, and weight sync; you write a TOML config and run a CLI command.
- Up to 2 concurrent runs, 256 rollouts per batch: Capacity is limited and allocated while reserved compute lasts.
Reinforcement learning fine-tuning for frontier-class coding models used to mean wrangling GPU clusters, writing custom trainers, and burning through cloud budgets. Prime Intellect Lab is trying to change that -- and this month it is sweetening the deal with something hard to ignore: free training runs on Poolside's Laguna XS.2, one of the strongest open-weight agentic coding models available right now.
What just landed
Prime Intellect has partnered with Poolside to make Laguna XS.2 available on its Hosted Training platform at no cost for a limited window, on a first-come, first-served basis while reserved capacity lasts. You get up to two concurrent training runs, each with up to 256 rollouts per batch. The goal is simple: let you specialize a top-tier coding model on your own environments without touching a single GPU config.
Laguna XS.2 is Poolside's first open-weight model -- a second-generation MoE built on everything learned from training the larger M.1. At 33B total parameters with only 3B activated per token, it is compact enough to run on a Mac with 36 GB of RAM via Ollama. That efficiency-to-capability ratio is exactly what makes it an interesting target for RL fine-tuning: you get a strong base without the compute overhead of a dense model.
The model under the hood
Laguna XS.2 scores 68.2% on SWE-bench Verified, 62.4% on SWE-bench Multilingual, 44.5% on SWE-bench Pro, and 30.1% on Terminal-Bench 2.0. SWE-bench Verified is the standard leaderboard for measuring how well a model can autonomously resolve real GitHub issues -- so 68.2% puts XS.2 firmly in the top tier for its size class.
The architecture is purpose-built for long-horizon agentic work. XS.2 uses sigmoid gating with per-layer rotary scales, enabling a mixed Sliding Window Attention (SWA) and global attention layout in a 3:1 ratio across 40 total layers. Sliding Window Attention limits each token's attention to a local window of 512 tokens rather than the full sequence, dramatically cutting KV cache memory. In practice, this means the model can handle long agent trajectories -- multi-step tool calls, large codebases -- without the memory blowup that plagues dense models at this context length.
Poolside replaced AdamW with a distributed implementation of the Muon optimizer, achieving the same training loss in roughly 15% fewer steps, with lower memory requirements -- only one state per parameter instead of two. The entire model was trained from scratch inside Poolside's internal Model Factory. XS.2 started pre-training just five weeks before its release, and shipped fully post-trained.
The weights are available under Apache 2.0 -- no royalties, no usage restrictions, fully commercial. This is a stark contrast to the more restrictive licenses used by some other labs.
The platform doing the heavy lifting
Prime Intellect Lab is a training platform for self-improving agents that connects environments, evaluations, hosted RL training, adapter deployments, and inference into one loop. The key abstraction is the environment -- a self-contained package that includes your dataset, a model harness (tools, sandboxes, context management), and a reward function that defines what success looks like.
Hosted Training runs large-scale reinforcement learning inside your environments. Prime Intellect manages the trainer nodes, rollout inference, orchestration, scheduling, auto-scaling, and weight sync. You write a small TOML config and fire a CLI command:
model = "poolside/Laguna-XS.2"
max_steps = 200
batch_size = 256
rollouts_per_example = 16
[[env]]
id = "your-org/your-coding-env"Lab spins up a dedicated orchestrator for your run, drives rollouts through your environment, scores trajectories with your rubric, and trains a LoRA adapter with the open-source prime-rl trainer. Inference workers update as training progresses, so rollouts always come from the latest weights. LoRA (Low-Rank Adaptation) is a technique that fine-tunes only a small set of additional weight matrices rather than the full model, keeping compute costs manageable.
You pay for the tokens that actually move the model -- not for GPU time you reserved and didn't use. During the Laguna promotion, even that token cost is waived.
Getting started in four steps
- Install the
primeCLI:uv tool install -U primeand runprime login - Set up your workspace:
prime lab setup-- this scaffolds anAGENTS.md, example environments, and starter configs - Pick or build an environment from the Environments Hub, or write your own using the
verifierslibrary - Launch:
prime train run configs/rl/your-config.tomland watch reward curves in the dashboard
Once the run finishes, you deploy the trained LoRA adapter through Prime Inference and hit it via an OpenAI-compatible API endpoint -- no changes to your existing application code.
Why this matters beyond the free tier
The deeper story here is about who gets to do RL post-training. Until recently, running large-scale agentic RL -- the kind that produced models like Laguna XS.2 itself -- required infrastructure that only well-funded labs could afford. Through research like INTELLECT-3 and thousands of hosted training runs, Prime Intellect's goal is to democratize access to a state-of-the-art agentic RL stack that was previously accessible only to the largest labs.
Prime Intellect believes RL is not only the path to AGI, but also the foundation for building AI-native products. The most successful future startups will emerge by creating novel and differentiated environments tailored to their needs. That thesis is now testable for free, on a model that already beats most closed alternatives in its size class.
The practical use-cases are broad:
- Domain-specific code agents -- train Laguna to resolve issues in your internal codebase, following your conventions and tooling
- Multilingual software engineering -- XS.2 already scores 62.4% on SWE-bench Multilingual; RL on your target language stack can push that further
- Custom tool use -- define an environment where the agent must call your internal APIs, and reward it for correct outcomes
- Research baselines -- run ablations on RL algorithms, reward shaping, or environment design without provisioning your own cluster
The capacity is limited and the window is short. If you have been waiting for a low-friction on-ramp to agentic RL fine-tuning on a genuinely capable open model, this is the clearest one yet.