Built with mlx-optiq, the MLX-native toolkit to quantize, fine-tune, and serve LLMs locally on Apple Silicon, no PyTorch and no cloud. Try the Lab · All OptiQ quants · Docs
A mixed-precision MLX quant of poolside/Laguna-XS-2.1, a sparse mixture-of-experts reasoning model built for coding and agentic work. Per-layer bit-widths come from a KL-divergence sensitivity pass on a six-domain calibration mix (prose · reasoning · code · agent · tool-call · constraint-bearing instructions): sensitive layers get more bits, robust ones fewer, at ~4.5 bits-per-weight average (20 GB on disk).
Stock mlx-lm has no laguna class. OptiQ ships a vendored, mlx-native port of the Laguna architecture (sigmoid MoE with a shared expert, QK-norm, GLM partial-rotary, softplus attention gate, hybrid full/sliding attention with dual RoPE) that registers with mlx-lm on import optiq. Install and import it before loading:
pip install "mlx-optiq>=0.4.7"
import optiq # registers the laguna arch with mlx-lm
from mlx_lm import load, generate
model, tok = load("mlx-community/Laguna-XS-2.1-OptiQ-4bit")
prompt = tok.apply_chat_template(
[{"role": "user", "content": "Write a Python function that returns the nth Fibonacci number."}],
tokenize=False, add_generation_prompt=True,
)
print(generate(model, tok, prompt=prompt, max_tokens=400))
Laguna is a reasoning model. It thinks before answering, so give it a generous max_tokens. For an OpenAI + Anthropic-compatible server with mixed-precision KV cache, tool-call healing, and prompt caching, use optiq serve:
Following the naming llama.cpp uses for its mixed quants (Q4_K_M and friends), the "4bit" label denotes the family, not the weighted average. The mixed allocation is what preserves capability at this size.
Six-metric Capability Score (the unweighted mean of MMLU, GSM8K, IFEval, BFCL, HumanEval, and HashHop). Scored in reasoning mode (generative MMLU + a large think-token budget), the setting that measures an always-on thinking model fairly.
Metric
OptiQ-4bit
MMLU (reasoning, 969 samples)
86.2%
GSM8K (1000 samples)
95.6%
IFEval (full set, strict)
76.5%
BFCL-V3 simple (200 calls)
89.0%
HumanEval (164 problems, pass@1)
82.9%
HashHop (long-context retrieval)
84.0%
Capability Score (mean of 6)
85.70
Every metric gets one equal vote. See the eval-framework writeup for the full methodology.
import os
os.environ["HF_ENDPOINT"] = "https://freeweights.allternit.com"
from transformers import AutoModelForCausalLM, AutoTokenizer
tok = AutoTokenizer.from_pretrained("mlx-community/Laguna-XS-2.1-OptiQ-4bit", revision="92398949b86bced5240a19a8f815afcb22c66d17")
model = AutoModelForCausalLM.from_pretrained("mlx-community/Laguna-XS-2.1-OptiQ-4bit", revision="92398949b86bced5240a19a8f815afcb22c66d17", device_map="auto")
mlx-community <huggingface.co/mlx-community>Newsgroups: alt.freeweights.textSubject: 
