What is Laguna S 2.1?
Laguna S 2.1 is Poolside's current flagship coding model, released on 21 July 2026 with weights on Hugging Face the same day. It is a mixture-of-experts model with 118 billion parameters, about 8 billion active per token, and a context of 1,048,576 tokens. It is text only, built for agentic coding, and thinks between tool calls; thinking can be switched off per request. It sits above Laguna XS 2.1 (33B) and, by Poolside's account, beats the larger Laguna M.1 (225B) at half the running size.
The weights are under OpenMDW-1.1, a permissive licence that allows commercial use and modification without asking. Poolside ships BF16, FP8, NVFP4, INT4 and GGUF versions plus a DFlash draft model for faster decoding. The BF16 checkpoint is about 236 GB, and Poolside says the model fits on a single NVIDIA DGX Spark. On Hugging Face it has 1,077 likes and 584,050 downloads in the last 30 days.
At launch Poolside listed a paid 1M-context OpenRouter endpoint at $0.10 per million input tokens, $0.20 per million output tokens and $0.01 for cache reads, plus a free 256K-context endpoint. It is also on Baseten, Vercel AI Gateway, Ollama and a free web chat.
How much does Laguna S 2.1 cost?
| Plan | Price | What you get |
|---|---|---|
| Weights | Free | OpenMDW-1.1 |
| OpenRouter (paid) | $0.10 in / $0.20 out per 1M tokens | $0.01 cache read, 1M context |
| OpenRouter (free) | Free | 256K context |
Open weights under OpenMDW-1.1. Launch OpenRouter price: $0.10 input / $0.20 output / $0.01 cache read per 1M tokens; free 256K-context endpoint.
Is Laguna S 2.1 worth it?
Laguna S 2.1 does far more agentic coding than its 8 billion active parameters suggest. Poolside reports 70.2% on Terminal-Bench 2.1 and 59.4% on SWE-Bench Pro, against 64.0% and 55.4% for DeepSeek-V4-Pro, which has six times the active parameters. On DeepSWE, the hardest long task Poolside ran, it scores 40.4% to DeepSeek-V4-Pro's 9.0%. Every evaluation run is published at trajectories.poolside.ai, which is more than most labs offer.
It is not the strongest open coder. Kimi K3 scores 88.3% on Terminal-Bench 2.1 and 69% on DeepSWE in the same table, and Poolside's numbers are its own. We found no Artificial Analysis Intelligence Index score for it, and Poolside's DeepSWE result came from its own harness rather than the leaderboard's.
The catch: Poolside lists the failure modes itself. It sometimes follows its memory of a tool rather than the definition a third-party harness gives it, writes broken JSON when a tool expects an array, and overthinks maths. Thinking is either off or at maximum, and with thinking off Terminal-Bench drops to 60.4%. The free OpenRouter endpoint may use your prompts for training, so private code belongs on the paid endpoint or your own GPUs.
Our take: Laguna S 2.1 is the open model to hand the routine half of a coding agent's work. At $0.20 per million output tokens a long agent session costs pennies, and when it stalls, GLM-5.3 or Kimi K3 can take the hard tickets.
Where it earns its place
A platform team running coding agents on every pull request
Fix failing tests and small bugs without paying frontier rates
Route routine tickets to Laguna S 2.1 with thinking on and preserved reasoning. At $0.10/$0.20 per million tokens, 1,000 sessions of 500K input and 20K output tokens cost about $54 before caching.
A regulated company that cannot send code to an outside API
Run an agentic coding model on its own hardware
Serve the FP8 or NVFP4 checkpoint with vLLM, SGLang or TRT-LLM on four GPUs, or the INT4 build on a DGX Spark. The OpenMDW-1.1 licence needs no permission for commercial use.
A developer comparing local coding models
Get an agent that runs in Ollama on a high-end desktop
ollama run laguna-s-2.1 pulls a Q4_K_M build with the chat template baked in. Hacker News users report 15 to 28 tokens a second locally, fine for chat but slow for interactive agent loops.
Before you commit
- Setup
- Download from Hugging Face, ollama run laguna-s-2.1, or call it through OpenRouter, Baseten or Vercel AI Gateway. vLLM needs the poolside_v1 tool and reasoning parsers; llama.cpp support is in Poolside's fork while upstream review continues.
- Integrations
- OpenAI-compatible APIs; vLLM, SGLang, TRT-LLM, Transformers and llama.cpp; agent harnesses including Kilo, Cline, OpenCode, OpenClaw, Hermes Agent and Poolside's pool.
- Data and privacy
- Self-hosting keeps all code on your hardware. OpenRouter's free endpoint may use inputs and outputs for training; the paid endpoint and dedicated Baseten deployments are the options for private code.
- Cost at scale
- $0.10/$0.20 per million tokens at launch with $0.01 cache reads; BF16 weights are about 236 GB and Poolside's vLLM recipe uses four GPUs.
- Limits
- 1,048,576-token context on the paid endpoint, 256K on the free one. Thinking is off or maximum only, and Poolside warns of invalid JSON for array arguments and occasional overthinking.
What’s good
- 70.2% on Terminal-Bench 2.1 with only 8B active parameters, per Poolside
- Permissive OpenMDW-1.1 licence with no revenue or user limits
- 1M-token context, with or without thinking
- Published evaluation trajectories and official FP8, NVFP4, INT4 and GGUF builds
What’s not
- Trails Kimi K3 and GLM-5.3 on the hardest agent tasks
- Benchmarks are self-reported; no independent index score yet
- Known tool-calling slips in third-party harnesses
- Thinking is off or maximum only, with no lower effort settings
Best for
Teams that want a cheap, permissively licensed open coding model they can run on a single workstation or call for cents.
Laguna S 2.1 alternatives
Laguna S 2.1: common questions
- Is Laguna S 2.1 worth it?
- Near-frontier agentic coding from a model small enough for one workstation, at a price that rounds to zero. We rate it 4.1 out of 5.
- How much does Laguna S 2.1 cost?
- Open weights under OpenMDW-1.1. Launch OpenRouter price: $0.10 input / $0.20 output / $0.01 cache read per 1M tokens; free 256K-context endpoint.
- What are the downsides of Laguna S 2.1?
- Trails Kimi K3 and GLM-5.3 on the hardest agent tasks; benchmarks are self-reported; no independent index score yet; known tool-calling slips in third-party harnesses; thinking is off or maximum only, with no lower effort settings.
- Who is Laguna S 2.1 best for?
- Teams that want a cheap, permissively licensed open coding model they can run on a single workstation or call for cents.
- How much does Laguna S 2.1 cost for a team or in production?
- $0.10/$0.20 per million tokens at launch with $0.01 cache reads; BF16 weights are about 236 GB and Poolside's vLLM recipe uses four GPUs.
- Is Laguna S 2.1 safe for company data?
- Self-hosting keeps all code on your hardware. OpenRouter's free endpoint may use inputs and outputs for training; the paid endpoint and dedicated Baseten deployments are the options for private code.
- What are the best Laguna S 2.1 alternatives?
- The closest alternatives we have reviewed are GLM-5.3, Kimi K3, Devstral 2 and DeepSeek-V4.1-Flash.
Answers checked Nov 7, 2026.
Make Laguna S 2.1? Show this rating on your site
