DEV.co
Open-Source LLM · poolside

Laguna-XS.2

Laguna-XS.2 is a 33B-parameter Mixture-of-Experts model with 3B activated parameters per token, designed for code-generation and agentic tasks that run locally. It uses sliding window attention for efficiency, supports long contexts (262k tokens), includes native reasoning/thinking modes, and is available under Apache 2.0 license. Benchmark results show competitive performance on SWE-bench and coding tasks.

Source: HuggingFace — huggingface.co/poolside/Laguna-XS.2
33.4B
Parameters
apache-2.0
License (OSI-approved)
Unknown
Context (tokens)
85.6k
Downloads (30d)

Key facts

Objective fields from the source. Values we can't verify are shown as “Unknown” rather than guessed.

FieldValue
Developerpoolside
Parameters33.4B
Context windowUnknown
Licenseapache-2.0 — OSI-approved
Modality / tasktext-generation
Gated on HuggingFaceNo
Downloads85.6k
Likes317
Last updated2026-07-01
Sourcepoolside/Laguna-XS.2

What Laguna-XS.2 is

33B total / 3B activated MoE model with 256 experts (1 shared), 40 layers (10 global attention, 30 sliding-window), FP8 KV cache quantization, 262k token context, Muon optimizer. Supports vLLM, SGLang, Transformers, TRT-LLM, Ollama, and mlx-lm. Includes speculative decoding via DFlash draft model. Trained with post-training and async off-policy agent RL.

Quickstart

Run Laguna-XS.2 locally

Load the open weights with 🤗 Transformers and generate — the same model, self-hosted.

quickstart.pypython
from transformers import pipelinepipe = pipeline("text-generation", model="poolside/Laguna-XS.2")out = pipe("Explain retrieval-augmented generation in one sentence.",           max_new_tokens=128)print(out[0]["generated_text"])

Swap in vLLM or Ollama for production-grade serving. DEV.co can stand up the inference stack.

Deployment

How you'd run it

A typical self-hosted path — open weights, an inference server, your application.

DEV.co builds each layer — from GPU infrastructure to the application.

Best use cases

Local code-generation and debugging

Optimized for software engineering tasks (SWE-bench Verified 69.9%) with sliding-window attention and low KV cache, runs on consumer hardware like Mac with 36GB RAM.

Agentic coding workflows

Native tool-calling support via poolside_v1 parser and interleaved reasoning enable multi-step coding tasks without external API dependencies.

Long-document code analysis

262k token context window supports large codebases and multi-file analysis with reasonable latency due to MoE efficiency.

Running & fine-tuning it

ESTIMATE: ~20–28 GB VRAM for full BF16 inference (3B active per token, 33B total with quantized KV cache in FP8). Documented as runnable on Mac with 36GB RAM; likely requires bfloat16 precision or lower for broader GPU compatibility. Exact precision-to-VRAM mapping not provided; verify with vLLM/TRT-LLM memory profiler.

Card does not document LoRA, QLoRA, or supervised fine-tuning feasibility. MoE architectures require specialized tooling for LoRA. Recommend: (1) consult poolside technical report or GitHub for fine-tuning guidance, (2) test adapter approaches on smaller MoE models (e.g., Mixtral 7B) first, (3) check vLLM/Transformers feature parity for LoRA injection on MoE.

When to avoid it — and what to weigh

  • Requires absolute state-of-the-art general language capability — Benchmarks show Laguna-XS.2 trails Qwen3.6-35B-A3B (73.4% vs 69.9% SWE-bench) and GPT-5.4 Nano (52.4% vs 46.3% SWE-bench Pro). Specialized for coding, not general-purpose tasks.
  • Need cloud-only solution with no local deployment — Model is optimized for and intended for local/on-prem deployment. If full managed service is non-negotiable, poolside offers API access but with different operational model.
  • Multi-modal input requirement — Model is text-to-text only; does not support images, audio, or other modalities.
  • Unknown deployment scale or latency SLAs — Speculative decoding (DFlash) adds complexity. Baseline inference latency and throughput not documented; requires testing in target environment.

License & commercial use

Apache 2.0 license. Permissive OSI-compliant license permitting use, modification, and distribution for commercial and non-commercial purposes.

Apache 2.0 is permissive and explicitly allows commercial use without attribution requirement. No gating, no API key restriction on model artifacts themselves. Free API access on poolside platform noted as time-limited; commercial deployment should self-host or negotiate commercial terms with poolside for ongoing API usage.

DEV.co evaluation signals

Editorial assessment — not user reviews. Directional, with an explicit confidence level.

SignalAssessment
MaintenanceActive
DocumentationStrong
License clarityClear
Deployment complexityModerate
DEV.co fitStrong
Assessment confidenceHigh
Security considerations

No explicit security audit, threat model, or vulnerability disclosure process documented. Card notes 'custom_code' in tags (likely for tool-calling/reasoning parsing); review remote-code-execution risks when using `--trust-remote-code` flag in vLLM. No data poisoning or adversarial robustness claims. Recommend: (1) audit poolside_v1 parsers before production, (2) sandbox tool execution when used as agentic model, (3) monitor poolside GitHub for security advisories.

Alternatives to consider

Qwen3.6-35B-A3B

Larger MoE (35B), stronger SWE-bench scores (73.4% Verified, 67.2% Multilingual, 49.5% Pro), higher Terminal-Bench (51.5%). Trade-off: slightly higher VRAM, less optimized for local Mac deployment.

Claude Haiku 4.5 (via API)

Proprietary, closed-weights alternative with comparable SWE-bench Verified (73.3%) and stronger Terminal-Bench (29.8% vs 35.7%). Advantages: fully managed, mature safety/security posture. Disadvantages: no local deployment, API-only, vendor lock-in.

Devstral Small 2 (24B dense)

Lightweight dense alternative (no MoE complexity) with strong SWE-bench Verified (68.0%). Trade-off: lower absolute performance, but simpler deployment and lower memory overhead if local VRAM is severely constrained.

Software development agency

Ship Laguna-XS.2 with senior software developers

Download the model from HuggingFace, self-host via vLLM or Ollama, and start building agentic code workflows without external API dependencies. For guidance on integration with your stack, consult our AI development services.

Talk to DEV.co

Related open-source tools

Surfaced by semantic similarity across the DEV.co open-source index.

Laguna-XS.2 FAQ

Can I use Laguna-XS.2 commercially without paying poolside?
Yes. The model weights are Apache 2.0 licensed, which permits commercial use. Download from HuggingFace and self-host (vLLM, Transformers, etc.) at no cost. However, if using poolside's API, confirm current terms; they offer free access for a limited time. Long-term commercial API use requires negotiation.
How much GPU VRAM do I need to run Laguna-XS.2 locally?
Estimated 20–28 GB VRAM in bfloat16 or lower precision. Documented as runnable on Mac with 36GB RAM. Exact memory depends on inference framework and quantization; test with vLLM or mlx-lm memory profiler in your target environment.
Does Laguna-XS.2 support fine-tuning or LoRA?
Not explicitly documented in the model card. MoE models require specialized LoRA tooling. Consult poolside's technical report or GitHub for guidance. Recommend testing fine-tuning feasibility with poolside before committing to production training pipelines.
What is the difference between Laguna-XS.2 and Laguna M.1?
Card mentions Laguna M.1 (225B) as a larger model; no side-by-side comparison provided. Infer: M.1 is for higher-quality, more complex tasks at cost of larger deployment footprint. XS.2 is optimized for local, resource-constrained agentic coding.

Custom software development services

From first prototype to production, DEV.co delivers software development services around tools like Laguna-XS.2. Our software development agency staffs experienced software developers and web developers for custom software development, web development, integrations, and ongoing support across open-source llms and beyond.

Deploy Laguna-XS.2 in Your Environment

Download the model from HuggingFace, self-host via vLLM or Ollama, and start building agentic code workflows without external API dependencies. For guidance on integration with your stack, consult our AI development services.