LocalClaw/LLM/Laguna XS 2.1
Open-weight MoE

Laguna XS 2.1

Poolside agentic coding MoE with 262K context, 33B total / 3B active parameters, OpenMDW-1.1 licensing and official Q4_K_M GGUF plus Ollama availability for 36GB-class local machines.

64 GB workstation 36 GB RAM Q4_K_M Local coding agent
Run with LocalClaw Compare all models
Parameters
33B (3B active, MoE)
Minimum RAM
36 GB
Model size
20.3 GB
Quantization
Q4_K_M

Can Laguna XS 2.1 run locally?

Laguna XS 2.1 is best for 64 GB workstations and larger Apple Silicon or NVIDIA setups.

Search for laguna-xs-2.1 in LM Studio or another GGUF-compatible runtime.

Model sourcepoolside/Laguna-XS-2.1-GGUF
chatcodereasoningagentpowerlong-context

Install path

01
Check RAM fitMinimum 36 GB RAM. Start with the Q4_K_M quant.
02
Load the modelSearch laguna-xs-2.1 in LM Studio.
03
Control locallyUse LocalClaw to manage models, agents, chat, channels and scheduled OpenClaw work.

Strengths

  • Official Poolside base weights and official GGUF conversion
  • Designed for local agentic coding and long-horizon terminal-style workflows
  • 33B total parameters with only 3B active per token keeps active compute compact
  • 262K context window with native interleaved reasoning support
  • Q4_K_M GGUF is about 20.3GB and Poolside targets 36GB-class Macs
  • Available through Ollama plus llama.cpp-compatible GGUF artifacts

Limitations

  • OpenMDW-1.1 is a custom license, not Apache 2.0 or MIT
  • Laguna architecture support is new and may require recent runtimes or patched llama.cpp builds
  • Poolside notes current macOS Metal chat behavior may need workarounds in some Ollama paths
  • Best quality and benchmark claims are still mostly vendor-reported
  • Long context can push real memory use far above the 20.3GB Q4_K_M file size

Best use cases

  • Local coding agent
  • Terminal task automation
  • Long-context repository analysis
  • Reasoning-heavy software engineering prompts
  • Private workstation code review
  • Ollama-based local assistant experiments

Capability profile

speed
6
quality
8
coding
9
reasoning
8

Technical notes

Developer
Poolside
License
OpenMDW-1.1
Context window
262,144 tokens
Architecture
33B total parameter Mixture-of-Experts language model with 3B activated parameters per token, 256 experts plus one shared expert, and mixed sliding-window plus global attention layers.

This model fits these next steps

Hardware fit is based on LocalClaw's RAM tier, model size and quantization metadata. Always leave memory headroom for your OS and runtime.

Workstation fitMac Studio M4 Max 64GBCUDA workstation classNVIDIA GB10 / DGX SparkLarge local models64GB RAM guide

Similar models to compare

North Mini Code 1.0 30B (3B active, MoE)Qwen AgentWorld 35B-A3B 35B (3B active, MoE)Bonsai 27B 27.3B (ternary / 1-bit)GPT-OSS (20B) 20B

Where to go next