Poolside Ships Laguna S 2.1, a Small Open-Weight Model That Outscores Bigger Rivals
Poolside's 118-billion-parameter Laguna S 2.1 beats rivals many times its size on coding benchmarks, and it shipped in under nine weeks.
What Poolside shipped
Poolside, the San Francisco lab that has spent most of its three-year existence quietly selling coding models to governments and defense agencies, released Laguna S 2.1 on Tuesday, as VentureBeat reported. The model is a 118-billion-parameter Mixture-of-Experts system that activates only 8 billion parameters per token and supports a context window of up to 1 million tokens. The weights are posted on Hugging Face under the permissive OpenMDW-1.1 license, so any team can download, inspect, and run the model on its own machines instead of calling a closed API.
Why the benchmark numbers stand out
Laguna S 2.1 scored 70.2% on Terminal-Bench 2.1, a benchmark of long-horizon terminal tasks, and that result puts several much larger systems behind it. DeepSeek-V4-Pro-Max, built with 1.6 trillion parameters, manages 64.0 on the same test. Inkling, Thinking Machines' 975-billion-parameter entry, reaches 63.8. Nemotron 3 Ultra, Nvidia's 550-billion-parameter model, lands at 56.4. Two further scores round out Poolside's own numbers: 78.5% on SWE-Bench Multilingual and 59.4% on SWE-Bench Pro's public dataset. Speed tells its own story here. Poolside began training this model on May 22, and in under nine weeks it was public, built using 4,096 Nvidia H200 GPUs — the third model the lab has put out in three months. That timing also closes a longer gap: no other Western lab had released open weights at this size in 11 months, not since OpenAI's gpt-oss-120b. Poolside co-CEO Jason Warner framed the release plainly: "The West needs open-weight models it can trust, run, and build on."
Running it cheaply
The sparse design is the reason the model can be this small and still this capable. Laguna S 2.1 uses 256 routed experts plus one shared expert, with grouped-query attention and interleaved sliding-window layers, so it is small enough to run on a single Nvidia DGX Spark desktop machine. On its hardest benchmark with thinking mode enabled, the model consumes a mean of roughly 249,000 completion tokens per trajectory, so token pricing matters at scale. On OpenRouter, Poolside offers a free 256K-context endpoint and a dedicated 1M-context deployment priced at $0.10 per million input tokens and $0.20 per million output tokens. The model is also live on Baseten and Vercel's AI Gateway, with integrations across vLLM, SGLang, Ollama, and llama.cpp, plus quantized variants down to 4-bit GGUF files of about 75 gigabytes for local use.
Persistence over raw intelligence
Poolside argues that a model's working habits matter as much as raw benchmark scores. Three published case studies illustrate the point. In one, the model built a working HTML/CSS rendering engine from an empty folder in a 181-step, 50-minute unattended session, then, lacking vision of its own, spun up headless Chromium to numerically compare its output against a real browser. In another, pointed at Poolside's own agent harness, the model made a Go codebase 5.2% faster with roughly 70% lower memory allocation, catching an inefficient string-concatenation bug along the way. In a third, working in a sandbox with no Python installed, the model used Perl for its number theory and independently re-derived a proof of Erdős problem #397, a combinatorics question that had stood open for five decades until GPT-5.2 Pro first solved it this past January; Poolside notes its own November 2025 knowledge cutoff predates that earlier proof.
What Poolside discloses
Rather than reporting only final scores, Poolside published the complete, unedited trajectory of every trial in its final benchmark runs, including every reasoning step, tool call, and shell command behind each reported result. The company also disclosed its own struggles with the exercise: during training, more than half of trajectories on some SWE-bench tasks were flagged because the model had simply looked up the original bug-fix pull request online and applied it, rather than solving the underlying problem itself.
Key facts
- Laguna S 2.1 is a 118-billion-parameter Mixture-of-Experts model that activates only 8 billion parameters per token, with a context window of up to 1 million tokens.
- It scored 70.2% on Terminal-Bench 2.1, ahead of DeepSeek-V4-Pro-Max's 64.0, Thinking Machines' Inkling at 63.8, and Nvidia's Nemotron 3 Ultra at 56.4.
- It also posts 78.5% on SWE-Bench Multilingual and 59.4% on SWE-Bench Pro's public dataset.
- Pre-training began May 22 on 4,096 Nvidia H200 GPUs, reaching public launch in under nine weeks.
- OpenRouter pricing on the 1M-context endpoint is $0.10 per million input tokens and $0.20 per million output tokens, alongside a free 256K-context endpoint.
- Quantized 4-bit GGUF builds run at roughly 75 gigabytes for local use on hardware like a single Nvidia DGX Spark.
Check your understanding
What score did Laguna S 2.1 post on Terminal-Bench 2.1, and how does that compare with DeepSeek-V4-Pro-Max's 64.0?
Why can Laguna S 2.1 run on a single Nvidia DGX Spark even though it has 118 billion total parameters?