Laguna XS 2.1: Local Coding LLM With 256K Context and Open Licence

On 2 July 2026, Poolside AI published Laguna XS 2.1 to Hugging Face and the Ollama model library. Practitioners on X noticed within hours: TeksEdge posted that "Laguna XS 2.1 is a small [powerful] coding model" from Poolside with GGUF and Ollama support already live. For teams running local AI, the release combines meaningful benchmark gains, a more permissive licence, and quantisation options that fit hardware already present in most SMB offices.

What Laguna XS 2.1 Is

Laguna XS 2.1 is a Mixture-of-Experts (MoE) coding model developed by Poolside AI, an enterprise software engineering AI company. According to Poolside's release documentation, the model has:

• 33 billion total parameters, 3 billion activated per forward pass
• 256,000-token context window (up to 32,000 output tokens)
• OpenMDW-1.1 licence, which allows commercial deployment

The MoE architecture is what makes local deployment practical for teams without GPU clusters. A dense 33B model at BF16 precision demands roughly 66 GB of memory bandwidth during generation. Because Laguna XS 2.1 activates only 3B parameters per step, inference throughput behaves more like a compact 3B model while quality draws on the full 33B parameter space. According to Poolside AI, the model runs on a Mac with 36 GB of unified memory, a configuration that covers Mac Studio M4 Max, MacBook Pro M4 Max, and similar configurations.

What Changed From Laguna XS.2

Poolside released Laguna XS.2 in late April 2026, where it scored 68.2% on SWE-bench Verified, a benchmark measuring how reliably a coding model resolves real GitHub issues. Laguna XS 2.1 targets SWE-bench Multilingual and improves that score by 5.4 points to 63.1%, as reported by Poolside AI.

Multilingual coding benchmarks test correctness across Python, JavaScript, TypeScript, Go, Rust, and other languages that appear in typical European SMB codebases. A 5.4-point lift on multilingual evaluation means the model is measurably less likely to produce syntactically valid code that silently breaks in a non-Python context, which matters if your team works across several runtimes, as is common in web and backend development.

The licence also evolved. Laguna XS.2 shipped under Apache 2.0. Laguna XS 2.1 ships under OpenMDW-1.1, a newer open-model licence framework described by Poolside as better suited to enterprise commercial use cases, maintaining the permissive character of the previous release.

Deployment Options

Ollama (single developer or small team)

Ollama added Laguna XS 2.1 at launch. The model is available at ollama.com/library/laguna-xs-2.1. From a terminal:

ollama pull laguna-xs-2.1
ollama run laguna-xs-2.1

Ollama automatically selects Q4KM quantisation, which requires approximately 20-24 GB of memory. This leaves working headroom on a 36 GB Mac and fits within the 24 GB VRAM of an NVIDIA RTX 3090.

vLLM and SGLang (multi-user team server)

For deployments where several developers query the model simultaneously, vLLM and SGLang offer continuous batching that Ollama does not provide. Both are listed as supported serving frameworks by Poolside. A single Mac Studio or a Linux workstation with an RTX 3090 running vLLM can serve a team of 5-20 developers with acceptable latency on coding tasks. Configuration is straightforward using the standard vllm serve command with the Hugging Face model identifier.

GGUF (llama.cpp, LM Studio, Jan)

For maximum hardware flexibility, including Windows workstations and older Apple Silicon, Poolside released GGUF checkpoints at poolside/Laguna-XS-2.1-GGUF, including BF16 and Q4KM variants. These load in LM Studio, Jan, or llama.cpp with no additional configuration.

The DFlash Speculator

Alongside the base model, Poolside released DFlash: a small draft model designed for speculative decoding with Laguna XS 2.1. Speculative decoding works by having the small draft model propose a sequence of tokens; the large model then verifies the entire sequence in one forward pass. When the draft model is accurate, throughput increases substantially. According to Poolside's documentation, DFlash approximately doubles the achieved tokens per second when used with Laguna XS 2.1, a significant benefit for interactive coding assistance, where fast first-token latency and continuous output matter.

Privacy and GDPR

Coding assistants that route requests through cloud APIs send code to third-party servers. Depending on what your codebase contains, this may include proprietary algorithms, internal API credentials embedded in configuration files, or logic that processes customer data. Whether that constitutes personal data under GDPR is a legal question specific to your circumstances, but the risk profile changes fundamentally when inference runs entirely on your own infrastructure: no request leaves your network, no input token reaches Poolside, GitHub, or any cloud provider.

This is a recurring concern for European companies in regulated sectors, healthcare, legal, financial services, where data residency is not optional. A local Laguna XS 2.1 deployment satisfies data residency requirements by design.

EU AI Act Considerations

The EU AI Act's transparency obligations for general-purpose AI systems are easier to satisfy when the model and its version are fully under your control. With an open-weights model like Laguna XS 2.1, you can document exactly which model version is running, the training data category, and the licence terms, all requirements that a black-box cloud API makes difficult or impossible to fulfil independently. Based on our reading of the Act's Art. 50 provisions, self-hosted open-weight models sit in a more defensible compliance position for SMBs than opaque cloud services, provided you maintain version records and usage logs.

Practical Applications for a 10-100 Person Team

• IDE code completion via a local OpenAI-compatible endpoint (VS Code, JetBrains, Neovim all support configurable backends)
• Automated code review as a CI pipeline step: send pull request diffs to the local model and receive structured review comments
• Technical documentation generation from existing source files, using the 256K context to load an entire module in one pass
• Natural-language to SQL for internal analytics, with the full database schema loaded in context
• Migration script drafting when moving between framework versions or database systems
• Internal chatbot for engineering SOPs, grounded by retrieval-augmented generation on your own documentation

Hardware Quick Guide

For BAFA digital consulting grants in Germany or Kit Digital subsidies in Spain, a local Laguna XS 2.1 deployment on existing office hardware may qualify as part of a broader AI integration project. The licence and open-weights nature of the model simplify the documentation requirements for such grant applications.

Getting Started

The fastest path to a working local coding assistant:

1. Install Ollama from ollama.com
2. Run ollama pull laguna-xs-2.1 (approximately 20 GB download)
3. Point your IDE's AI plugin to http://localhost:11434/v1 with any model name
4. Test with a real task from your own codebase

If you want to discuss how Laguna XS 2.1 fits into your team's development workflow, or which hardware and serving setup makes sense for your team size, start a conversation with Freshlab. We help European SMBs design and implement local AI infrastructure that stays within GDPR boundaries and can scale without vendor lock-in.