A mixed-precision 4-bit MLX quantization of poolside/Laguna-XS-2.1 (revision c5f36269bbdbd3f27fddc9a9f9dbae0cf2cf57db): Poolside's coding-agent MoE (33B total / ~3B active per token, 40 layers, 262k context). Built for Midium.
oqe_code_multilingual, collected until every routed expert was seen — 286 samples × 512 tokens, all 29,952 experts covered).lm_head 8-bit; router gates and norms unquantized.config.json (including the corrected YaRN attention_factor), the chat template, tokenizer, generation config and Poolside's modeling_laguna.py / configuration_laguna.py (not needed by MLX runtimes).Weights: 19.5 GB (bf16 source: 66.9 GB).
Against the bf16 model — teacher-forced next-token loss with the text placed as an assistant reply in the chat template (Laguna is out of distribution on raw text); 6 × 2,048-token code windows and 4 prose windows, last 1,024 positions each:
| NLL change vs bf16 | Top-1 agrees with bf16 | |
|---|---|---|
| Code (6,144 tokens) | +0.0043 ± 0.0083 | 94.8% |
| Prose (4,096 tokens) | +0.0025 ± 0.0129 | 88.6% |
Speed (Midium engine, M5 Max 128 GB, greedy, ~1,300-token code reply): 124–128 tok/s plain decode; 250–252 tok/s with Poolside's DFlash drafter.
Runs in any MLX runtime that supports the laguna architecture (mlx-vlm ≥ 0.7, Midium). Recommended sampling from Poolside's card: temperature 1.0, top_k 20, top_p 1.0.
Same as the base model: OpenMDW-1.1 (see LICENSE).
model-00003-of-00004.safetensors 5.1 GBmodel-00002-of-00004.safetensors 5.1 GBmodel-00001-of-00004.safetensors 5.1 GBmodel-00004-of-00004.safetensors 4.1 GBpoolside/Laguna-XS-2.1