One-line: 3B-active MoE model with 70.9% SWE-bench Verified, designed for agentic coding and long-horizon tasks on local hardware.
Coding benchmarks
Benchmark
Score
Source
HumanEval
—
—
HumanEval+
—
—
MBPP
—
—
LiveCodeBench
—
—
Aider polyglot
—
—
SWE-bench Verified
70.9%
lab
SWE-bench Multilingual
63.1%
lab
SWE-bench Pro (Public)
47.6%
lab
Terminal-Bench 2.0
37.5%
lab
Agentic / tool-use benchmarks
Benchmark
Score
Source
BFCL
—
—
τ-Bench
—
—
ToolACE
—
—
GAIA
—
—
Reasoning (context only)
Benchmark
Score
Source
MMLU / MMLU-Pro
—
—
GPQA Diamond
—
—
MATH
—
—
AIME 2025
—
—
Efficiency
Active params: 3B (total 33B, 256 experts with 1 shared expert)
Context length: 262,144 tokens
VRAM (fp16 rough): ~6 GB active (3B params × 2 bytes), but full model requires ~66 GB for all experts; FP8 KV cache reduces memory per token
Architecture: MoE with mixed SWA (sliding window attention) and global attention layers in 3:1 ratio across 40 layers
Local-ready: Runs on Mac with 36 GB RAM; available on Ollama and llama.cpp
What makes it notable
Laguna XS 2.1 is an upgraded version of Laguna XS.2 with a +5.4% jump on SWE-bench Multilingual (57.7% → 63.1%) and stronger terminal-style task performance. It uses native reasoning support with interleaved thinking between tool calls, and the OpenMDW-1.1 license allows commercial and non-commercial use. The model is compact enough for local deployment (3B active params) while competing with much larger models: it beats Claude Haiku 4.5 on SWE-bench Verified (70.9% vs 73.3% is close) and significantly outperforms Qwen3.6-35B-A3B on Terminal-Bench 2.0 (37.5% vs 51.5% is actually worse, but the model is designed for different use cases). The mixed attention layout (SWA + global) and FP8 KV cache make it efficient for long-context agentic work.