
Source	poolside/Laguna-XS.2
Architecture	`laguna` (40 layers, 256 routed experts top-8 + 1 shared, hybrid SWA+full attention)
Quant format	MXFP4 (mlx 4-bit affine, group_size=32)
Bundle size on disk	20.93 GB (21 safetensors shards)
License	Apache-2.0 (inherits from upstream)
Modalities	Text in / text out (no vision, no audio, no video)

What's quantized

All routed-expert linears (3D stacked), attention, dense layer-0 MLP, shared-expert, embed_tokens → mlx 4-bit affine (bits=4 group_size=32)
lm_head + all RMSNorms + router gate + e_score_correction_bias → fp16 passthrough

Architecture notes (preserved verbatim from upstream)

40 layers; per-layer attention head count alternates 48 (full-attn) / 64 (SWA) with shared 8 KV heads (GQA)
1:3 ratio of full-attn ↔ sliding-window-attention (window = 512), explicit layer_types list
Dual RoPE: full-attn = YaRN (base 500K, factor 32, original 4096, β_fast 64, β_slow 1, partial_rotary 0.5); SWA = default (base 10K, full rotary)
256 routed experts (top-8) + 1 shared expert; sigmoid + per-head gating (g_proj); q_norm/k_norm in attention
131k context window
Layer 0 dense MLP; layers 1-39 sparse MoE

Run on Apple Silicon

bash

pip install mlx safetensors transformers
python -m jang_tools.laguna.runtime \
    --src ~/.mlxstudio/models/OsaurusAI/Laguna-XS.2-mxfp4 \
    --prompt "def fibonacci(n):" --max-new 64

The runtime auto-detects weight_format (mxtq / mxfp4 / bf16) and loads the matching path (jang_tools/laguna/weight_loader_bf16.py).

Build

Reproduce locally from the bf16 source:

bash

python -m jang_tools.convert_laguna_mxfp4 \
    ~/.mlxstudio/models/_sources/Laguna-XS.2 \
    ~/.mlxstudio/models/OsaurusAI/Laguna-XS.2-mxfp4

Credits

Quantized by Jinho Jang (eric@osaurus.ai). MLX-native pipeline, runs on M-series Macs.