A Coding-Model Lab, Not Another General API
Poolside is a Paris- and San Francisco-based AI lab that has spent the past two years building a single product line: the Laguna family of foundation models for software engineering. The lab does not market a general-purpose chat model. The July 2026 Laguna S 2.1, the June 2026 Laguna XS 2.1, and the April 2026 Laguna M.1 are all tuned for agentic coding, terminal-style tasks, and the multi-step verification loops that a real coding workflow runs through.
That focus shows up in three structural choices. First, every Laguna model is released as open weights under the OpenMDW-1.1 license, with BF16 plus FP8, INT4, NVFP4, GGUF, and MLX quantizations available on Hugging Face. Second, the API surface is OpenAI-compatible and exposed through OpenRouter as well as Poolside's own hosted endpoint, so existing OpenAI SDK clients work without code changes. Third, the docs.list six local inference paths (the poolside pool CLI, vLLM, SGLang, Ollama, llama.cpp, MLX, and NVIDIA TensorRT-LLM), so a single DGX Spark can host Laguna S 2.1 and a 36GB-RAM Mac can host Laguna XS 2.1.
This article covers the three-model lineup as it stands in September 2026, the verified per-million-token rates on OpenRouter, what the OpenMDW-1.1 license actually permits, where the agentic-coding benchmarks actually sit, and the trade-offs that come with a coding-specialized provider that does not publish a marketing pricing page.
The Three-Model Lineup
The Laguna family has three active models plus the open-weight release notes on Hugging Face. Each model targets a different deployment surface.
| Model | Released | Architecture | Context | Target surface |
|---|---|---|---|---|
| Laguna S 2.1 | July 2026 | 118B total / 8B active MoE | 1M tokens | Datacenter / DGX Spark single-node |
| Laguna XS 2.1 | June 2026 | 33B total / 3B active MoE | 256K tokens | Mac (36GB RAM) / single GPU |
| Laguna M.1 | April 2026 | Mid-tier agent | 256K tokens | Mid-scale coding agents |
Laguna S 2.1 is the flagship. It is a 118B-total / 8B-active Mixture-of-Experts model designed for long-horizon agent tasks. The 1M-token context window is what sets it apart from peer models in the same class — Anthropic's Sonnet 5 and OpenAI's GPT-6 both sit below that mark for first-tier models. The model supports native reasoning with two settings: thinking off or max thinking. The lower-effort controls (low, medium, high, xhigh) are not exposed in this release; if you need fine-grained cost control over reasoning, that is a structural limitation. Reasoning is on by default and surfaces through the OpenRouter reasoning object.
Laguna XS 2.1 is the small-footprint variant. The 33B-total / 3B-active MoE architecture fits a Mac with 36GB of RAM, and the KV cache is quantized to FP8 to keep memory-per-token low. The June 2026 release added native reasoning support with interleaved thinking between tool calls, plus measurable gains on SWE-bench Multilingual (+5.4% over Laguna XS.2) and Terminal-Bench 2.0 (+1.8% over XS.2). For a developer coding on a laptop, this is the most interesting entry point.
Laguna M.1 sits in the middle. Its 256K context and the benchmark profile (SWE-bench Verified 65.4%, Terminal-Bench 2.0 32.7%) position it for mid-scale coding agents that run on dedicated servers rather than laptops or full datacenter racks. Poolside describes M.1 as the most useful variant when the workflow is "explore an unfamiliar codebase, edit multiple files, run tests, iterate on a fix."
Verified Pricing on OpenRouter
Poolside does not publish a marketing pricing page. The verified rates below come from the OpenRouter model catalog and from the docs/release-notes/models.md release notes.
| Model | Input /1M | Output /1M | Cached input /1M | Context |
|---|---|---|---|---|
| Laguna S 2.1 | $0.09 | $0.18 | $0.009 | 1M |
| Laguna XS 2.1 | $0.06 | $0.12 | $0.003 | 256K |
| Laguna S 2.1 :free | $0 | $0 | — | 256K (rate-limited) |
| Laguna XS 2.1 :free | $0 | $0 | — | 256K (rate-limited) |
Two structural details matter. First, the cached-input ratio on both models is 10x cheaper than uncached input. On Laguna S 2.1, that means a system prompt or skill file at $0.009/M instead of $0.09/M; on Laguna XS 2.1, $0.003/M instead of $0.06/M. For an agent that carries a long system prompt across many turns, the cached tier is the single biggest lever on the bill.
Second, the free variants on OpenRouter are capped at 256K context, not the full 1M available on the paid Laguna S 2.1. A workload that needs the full 1M window must go through either Poolside's own hosted endpoint or OpenRouter's paid Laguna S 2.1. The chat.poolside.ai browser sandbox does not require an account but is for testing prompts, not production traffic.
For a budget comparison, Laguna S 2.1 at $0.09/$0.18 sits below Anthropic Sonnet 5 ($3/$15 list) and GPT-6 main-tier ($2.50/$10 list). It is comparable in input price to Qwen3.8-Flash and noticeably below GPT-6-Astra at $10/$50. The price-to-context ratio is where Laguna S 2.1 stands out: 1M context at $0.09/M is the best cost-per-token-of-context figure on OpenRouter for a 100B+ coding-specialized model.
OpenMDW-1.1: The Custom Open Weight License
All three active Laguna models are released under OpenMDW-1.1, a Poolside-authored license. The license text is published alongside each model card on Hugging Face, and the docs explicitly note that OpenMDW-1.1 is not a standard open-source license. The key terms:
- Commercial use: permitted for production deployments, including for products sold to third parties.
- Modification and redistribution: permitted with attribution; redistribution mechanics are spelled out in the license.
- Acceptable use: Poolside retains use-based restrictions (the standard set for AI safety, not a non-compete).
- Patent grant: included, mirroring Apache 2.0 style.
The practical impact is that a startup can ship a product built on Laguna S 2.1 weights, including modifying them and serving them via vLLM, without negotiating a separate license. The boundary conditions to verify with legal are: acceptable-use specifics (which use cases are explicitly out of scope), attribution mechanics when the weights are fine-tuned and redistributed, and any regional restrictions. The license is closer to Apache 2.0 than to a research-only license, but it is not identical.
If you are deploying on a customer's hardware (private cloud, on-prem, air-gapped), the OpenMDW-1.1 license is the right level of permissiveness. If you need a license that has been formally OSI-approved or that has been audited by a major open-source foundation, that is a gap.
Benchmarks: The Agentic Coding Profile
Poolside publishes benchmark numbers on its release notes pages, with full evaluation trajectories hosted at trajectories.poolside.ai. The benchmarks are heavy on coding and terminal tasks.
| Benchmark | Laguna M.1 | Laguna XS 2.1 | Notes |
|---|---|---|---|
| SWE-bench Verified | 65.4% | — | Up from 55.6% on Malibu 2.2 |
| SWE-bench Multilingual | 57.4% | 63.1% | XS 2.1 +5.4% over XS.2; M.1 up from 31.1% |
| Terminal-Bench 2.0 | 32.7% | 37.5% | XS 2.1 +1.8% over XS.2; M.1 up from 16.9% |
Two readings matter. First, the gains over predecessors are real and documented. Laguna M.1's jump on SWE-bench Multilingual from 31.1% to 57.4% in one generation is significant; on Terminal-Bench 2.0 it nearly doubles (16.9% to 32.7%). XS 2.1's smaller absolute numbers are still meaningful — +5.4% on SWE-bench Multilingual is a meaningful gain on a benchmark that resists easy progress.
Second, Poolside's coverage is narrower than Anthropic's or OpenAI's. Poolside publishes SWE-bench, Terminal-Bench, and internal coding traces; it does not publish AIME, GPQA, MMLU-Pro, or the broader academic reasoning benchmarks that frontier labs use to argue general capability. A buyer who needs Laguna to perform well on non-coding tasks is in lower-confidence territory and should run their own evaluation.
The full evaluation trajectories at trajectories.poolside.ai are an unusual plus — most vendors publish aggregate benchmark scores, not the per-step traces that show where the model fails or recovers. For a coding agent integration, this is useful for debugging prompts and tool-selection logic.
The API Surface
Both the Poolside-hosted endpoint and the OpenRouter relay expose an OpenAI-compatible API. Authentication is via Bearer token, and the model IDs are poolside/laguna-s-2.1, poolside/laguna-xs-2.1, and the :free variants on OpenRouter. Streaming is supported, and reasoning content is surfaced through OpenRouter's standard reasoning object.
For agent integration, the Poolside pool CLI is the canonical local control plane. It supports interactive mode, automated mode, and integration with Cline, Kilo Code, GitHub Actions, n8n, OpenClaw, Hermes, OpenCode, Pi, Goose, and other ACP-compatible agent frameworks. The Poolside docs.list integrations with JetBrains, Neovim, Zed, VS Code (via Copilot and Poolside Assistant), and several other IDE surfaces.
The Poolside-hosted endpoint supports the OpenAI Chat Completions schema; the Responses API path and tool-calling semantics follow the OpenAI standard. There is no separate MCP or function-calling protocol to learn. Existing OpenAI SDK clients work without code changes; the only delta is the model name.
The Trade-Offs to Weigh
Reasoning is binary. Laguna S 2.1 supports only thinking off or max thinking. There is no fine-grained reasoning.effort enum (xhigh / high / medium / low / minimal / none) that ACP clients support. If your cost model depends on stepping reasoning effort down for cheaper turns, this is a structural gap. Poolside is shipping the lower-effort modes in a later release but they are not available today.
License is custom, not OSI-approved. OpenMDW-1.1 permits commercial use, modification, and redistribution, but it is a Poolside-authored license. Some procurement processes require an OSI-approved license (Apache 2.0, MIT, BSD). The license text is published and auditable, but legal review is appropriate before any deployment that requires the procurement box to be checked.
Benchmark coverage is narrow. Poolside publishes coding benchmarks (SWE-bench, Terminal-Bench) and internal traces. It does not publish the broader academic reasoning benchmarks (MMLU, GPQA, AIME) that frontier labs use to demonstrate general capability. If Laguna is asked to handle non-coding tasks, expect to run your own evaluation rather than relying on published numbers.
No marketing pricing page. Poolside does not publish a pricing page on its main site. The verified rates come from OpenRouter and the docs/release-notes/models.md release notes. For a procurement team that wants a single source of truth, the absence of a marketing pricing page is a documentation gap that needs a workaround.
No mainland China endpoint. The Poolside-hosted endpoint and chat.poolside.ai both serve from non-CN infrastructure. A user in mainland China must proxy or self-host. OpenRouter's free variants are similarly regional.
The Bottom Line
Poolside is a coding-specialized AI lab with three active models under the Laguna name, all released as open weights under OpenMDW-1.1. The flagship is Laguna S 2.1 (118B/8B MoE, 1M context), priced at $0.09 input and $0.18 output per million tokens on OpenRouter, with cached input at $0.009/M (10x discount). Laguna XS 2.1 (33B/3B MoE, 256K context) is the small-footprint variant at $0.06/$0.12 with $0.003 cached input. Both have free variants on OpenRouter capped at 256K context.
The agentic-coding benchmark profile is real and improving — Laguna M.1's SWE-bench Verified 65.4%, SWE-bench Multilingual 57.4%, and Terminal-Bench 2.0 32.7% are competitive in the 100B-class coding-specialized model tier. The trade-offs are the ones the architecture implies: reasoning effort is binary, license is OpenMDW-1.1 (not OSI-approved), benchmark coverage is narrow, and there is no marketing pricing page.
The API is OpenAI-compatible and exposed through both Poolside's hosted endpoint and OpenRouter. Local inference is supported on vLLM, SGLang, Ollama, llama.cpp, MLX, and NVIDIA TensorRT-LLM. A single NVIDIA DGX Spark is sufficient to host Laguna S 2.1; a 36GB-RAM Mac is sufficient to host Laguna XS 2.1. For a team building an agentic coding workflow with a long-context requirement and a cost-per-token-of-context budget, Laguna S 2.1 at $0.09/M with 1M context is the standout number on OpenRouter today.
- Laguna S 2.1: $0.09 input / $0.18 output / $0.009 cached per 1M (10x cache discount); 1M context
- Laguna XS 2.1: $0.06 input / $0.12 output / $0.003 cached per 1M; 256K context; runs on 36GB-RAM Mac
- Laguna M.1: 256K context; SWE-bench Verified 65.4%, Terminal-Bench 2.0 32.7%
- OpenAI-compatible API; hosted on Poolside endpoint + OpenRouter
- Open weights under OpenMDW-1.1 (BF16 + FP8 / INT4 / NVFP4 / GGUF / MLX)
- Free variants on OpenRouter: Laguna S 2.1 :free and Laguna XS 2.1 :free (256K context, rate-limited)
- Local inference: vLLM, SGLang, Ollama, llama.cpp, MLX, NVIDIA TensorRT-LLM