Poolside Laguna M1 Free Acces: API, Weights, and Limits

Poolside laguna M1 free access Poolside laguna M1: yes, you can currently obtain Laguna M.1’s downloadable model weights at no charge under Apache 2.0, including the post-trained agentic-coding checkpoint and quantized variants. But do not read that as a promise of permanently free hosted API inference: Poolside’s April 2026 launch described API and OpenRouter access as free only “for a limited time,” while its current getting-started page advertises free Preview access for newer Laguna XS 2.1 and Laguna S 2.1 models rather than M.1.

Is Poolside Laguna M1 free?

The useful answer is “free weights, not necessarily free compute.” Laguna M.1 is publicly downloadable from Poolside’s official Hugging Face organization under the Apache 2.0 license. That gives developers broad permission to use, modify, and distribute the model subject to the license terms; it is a genuinely open-weight release, not merely a gated hosted-model trial.

The catch is that model access and model execution are different products. A 225B-total-parameter mixture-of-experts model still needs substantial storage, memory, infrastructure, and electricity to serve. The original BF16 repository is roughly 452 GB; Poolside also publishes FP8 and NVFP4 checkpoints, with the FP8 files listed at about 239 GB. Downloading the weights may be free, but operating them is not.

Poolside’s own release materials make this distinction fairly clear. Its April launch said both Laguna M.1 and Laguna XS.2 could be used for free for a limited time through Poolside’s API and OpenRouter. That wording was time-bounded. As of August 21, 2026, the current Poolside quick-start page explicitly markets free Preview use for Laguna XS 2.1 and Laguna S 2.1, so developers should not budget a production workflow around an assumed ongoing zero-cost M.1 endpoint.

How can I access Poolside Laguna M1?

There are three materially different access routes. Pick one based on whether you want to evaluate the model quickly, call it from an application, or control the full serving stack.

1. Download the weights from Poolside’s official Hugging Face model card. This is the clearest route for durable, no-cost access to the model artifact itself. The post-trained Laguna M.1 checkpoint is intended for chat, reasoning, tool calling, and agentic coding; do not accidentally choose Laguna M.1-base unless you specifically want a pre-training-only completion model.
2. Use Poolside’s agent and platform flow. The Laguna M.1 model card describes installing Poolside’s terminal agent, pool, and logging in to obtain a free API key. An API key is an account credential, however, not evidence that all future M.1 tokens are free or unlimited.
3. Use a third-party hosted route when it lists the model. Poolside originally named OpenRouter as an option, but provider catalogs and prices change. Check the exact model listing, per-token price, context limit, retention terms, and availability immediately before integrating it.

Are Laguna M1 weights open source or open weight?

For practical software use, Laguna M.1 is more than just a remotely callable model: its weights are released under Apache 2.0. The official model card labels the checkpoint Apache 2.0 and provides deployment instructions for vLLM, SGLang, and Transformers. That means a team can inspect the published configuration, download the checkpoint, select its own inference engine, and run it behind its own OpenAI-compatible endpoint.

Still, avoid turning “Apache 2.0 weights” into a sweeping claim that every part of the system is reproducible. Open weights describe availability and licensing of the released parameters. They do not automatically provide the original training data, full training pipeline, every internal evaluation environment, or free infrastructure. Poolside published a substantial technical report and the model configuration, which is valuable, but reproducing a 225B MoE training run is a separate question from downloading and serving the released artifact.

Can I run Poolside Laguna M1 locally?

Yes, but “locally” here usually means on infrastructure you control, not on a typical developer laptop. Laguna M.1 has 225.8B total parameters and activates 23.4B parameters per token. Poolside’s model card says that loading the BF16 checkpoint in Transformers requires substantial multi-GPU memory, and its SGLang serving example uses tensor parallelism across eight shards. The active-parameter count helps explain runtime compute, but the full model’s weights and KV cache still determine whether your hardware can actually host it.

Start with the quantized checkpoints if your purpose is experimentation rather than highest-fidelity inference. Poolside publishes FP8 and NVFP4 variants and states that SGLang detects their quantization configuration automatically. Before treating a quantized deployment as equivalent to the published headline benchmark, test the exact checkpoint, server version, context length, tool parser, sampling parameters, and coding-agent harness you will use. Agentic coding results are highly sensitive to the surrounding system, not only the base model.

What do I need to serve Laguna M1?

At minimum, plan for model storage, GPU memory across one or more compatible GPUs, a serving engine, and an API surface your tools can call. Poolside documents upstream support in vLLM, SGLang, Transformers, and TRT-LLM. For an application team, a sensible evaluation path is to serve the post-trained checkpoint behind an OpenAI-compatible endpoint, send a small repeatable suite of repository tasks, and capture cost, latency, tool-call reliability, and patch quality before committing to a larger deployment.

Use the post-trained Laguna M.1 rather than the base model for coding agents. Poolside describes Laguna M.1-base as a raw text-completion checkpoint without instruction following, reasoning, or tool-calling behavior. The regular Laguna M.1 release is the relevant checkpoint for an agent that needs to inspect files, propose edits, invoke tools, and continue over a long task.

Is a free API key the same as free M1 inference?

No. A free API key can mean no charge to create credentials, promotional credits, a preview quota, or access to a metered service. It does not by itself establish the price of every input and output token, the rate limits, the retention policy, or how long an offer will remain available. This is particularly important because Poolside’s own materials use both phrases: the M.1 model card refers to getting a free API key, while the launch announcement described hosted model use as free for a limited time.

For a one-off test, sign up and inspect the current platform terms before sending proprietary code. For a project, record the date, model ID, token pricing, request limits, and data-handling terms in the same engineering decision record where you would record a cloud dependency. “Free” is a useful way to lower the evaluation barrier; it is not a deployment architecture.

Should I use Laguna M1 or a newer Poolside model?

Use M.1 if you specifically want its 225B-A23B long-horizon coding model, want to experiment with the released Apache 2.0 checkpoint, or have infrastructure capable of serving it. But if your actual question is “what can I try free right now through Poolside’s supported preview workflow?”, the current answer points to Laguna XS 2.1 and Laguna S 2.1. Poolside’s getting-started page names those newer models for free Preview access in its terminal-based pool agent.

That is not a verdict that one model will solve every repository task better. It is a reminder to separate four decisions that are often collapsed into one: which checkpoint is most capable for the task, whether its weights are available, whether hosted inference is currently free, and whether your organization can afford to operate it. Laguna M.1 gives you unusually strong control over the second question because the weights are public; it does not eliminate the other three.

Try Cline when you are comparing coding-model access options

If this investigation is really about choosing a practical coding-agent workflow rather than collecting model files, Cline is an open-source AI coding assistant for VS Code that can read and write across files, run commands, and work through multi-step changes with developer approval. Its Plan and Act workflow is useful when you want to evaluate how a model behaves on real repository work instead of on a single chat prompt.

Cline is free for individual developers, while inference is usage-based: you can bring your own API keys or use its provider options. That does not change Poolside Laguna M.1’s availability or pricing terms, but it is relevant to the core access question: a model-agnostic coding-agent workflow lets you keep the harness and evaluation tasks stable while you compare the cost, capability, and control trade-offs of the models and endpoints you choose.