Laguna XS 2.1 (33B-A3B)
Agentic coding MoE, 33B total / 3B active, 262K context, OpenMDW-1.1 (fully permissive). Vendor says it runs on a 36 GB Mac.
Recommended quants
- Official Q4_K_M (20.3 GB)
poolside/Laguna-XS-2.1-GGUFOllama default tag. - Ollama q8_0 / bf16 (66.9 GB)
ollama laguna-xs-2.1 - APEX GGUF
mudler/Laguna-XS-2.1-APEX-GGUFPopular community quant; sizes not checked.
Chat template Caveats
Reasoning on by default; pass enable_thinking=False to skip <think>.
Fix: vLLM: use a build with PR #47311 or the poolside_v1 tool parser silently drops some tool calls.
Sampling settings
| Official | temp 1.0, top_k 20 |
|---|
Runtimes
| vLLM | Caveats | Needs vllm PR #47311 for reliable tool calls. |
|---|---|---|
| llama.cpp | Caveats | Needed PR #25165 at launch. macOS Metal: empty output bug until PR #25389. |
| Ollama | Caveats | Official library. Apple Silicon may return empty output (Metal f16 overflow) until llama.cpp #25389 ships. |
| LM Studio | Unverified | |
| MLX | Works | mlx-lm recommended by poolside. |
Tool calling in agent tools
| Claude Code (local endpoint) | Unverified | |
|---|---|---|
| OpenCode | Unverified | |
| Cline | Unverified | |
| Aider | Unverified |
Known issues
Sources checked
Spotted something wrong or a new fix? Corrections are reviewed before publishing. (Submission form coming soon.)