mlx-community/Laguna-S-2.1-OptiQ-2bit

Built with mlx-optiq, the MLX-native toolkit to quantize, fine-tune, and serve LLMs locally on Apple Silicon, no PyTorch and no cloud. All OptiQ quants · Docs · Laguna family

A 117-billion-parameter coding model that runs in 9.4 GB of RAM on a Mac. This is a 2-bit mixed-precision MLX quant of poolside's Laguna-S-2.1 (235 GB at bf16), produced by mlx-optiq. It is 41 GB on disk. While it generates, only ~9.4 GB sits in RAM: attention, the router, the shared expert and the layer edges stay resident, and the routed mixture-of-experts weights stream off the SSD as the router selects them.

Laguna-S is a sparse mixture-of-experts reasoning model built for coding and agents: 48 layers, 256 experts with 10 active per token. Asked to write Flappy Bird as a single HTML file, the 2-bit model produced a complete, working game with gravity, pipe collisions, scoring and restart, all in one pass. Here it is, playing the game it wrote:

The full game it generated is in this repo as flappy_bird.html. Open it in any browser.

What it is

Property	Value
Base	poolside/Laguna-S-2.1 (sparse MoE, 256 experts, 10 active per token, 48 layers)
Method	OptiQ static, structural per-layer bit allocation, no calibration
Bit-widths	4-bit on attention, router, shared expert and layer edges; 2-bit on the routed experts
Achieved bits-per-weight	3.01
On disk	41 GB
Resident while running	~9.4 GB (routed experts streamed)
Decode speed	~3 tok/s on an M3 Max, SSD-bound

No Capability Score is published for this quant. At 2-bit on the routed experts the point of the artifact is that a 117 B MoE runs at all on consumer Apple Silicon and stays coherent enough to write working code. For benchmarked quality in the Laguna family, use Laguna-XS-2.1-OptiQ-4bit (Capability Score 85.81).

For a model this large, exact calibration-driven sensitivity is impractical (it would run for days and needs the full model resident as a reference), so OptiQ's static method assigns bits from architecture alone. See the methods comparison.

Run it

Laguna ships under an architecture stock mlx-lm does not know, so import optiq once to register it:

pip install "mlx-optiq>=0.4.7"

The routed experts are too large to sit resident, so serve it with SSD expert streaming. optiq serve turns this on automatically for a MoE quant that would not fit in RAM (--stream-experts forces it):

optiq serve --model mlx-community/Laguna-S-2.1-OptiQ-2bit

That gives you an OpenAI + Anthropic-compatible endpoint with mixed-precision KV cache, tool-call healing and prompt caching. Only the routed experts stream per token; attention, the router and the shared expert stay resident, so the footprint stays ~9.4 GB no matter how large the model on disk is. Laguna is a reasoning model, it thinks before answering, so give it a generous token budget.

Notes

This is an extreme quant. 2-bit on the routed experts is lossy, and reference-quality generation should use the bf16 weights or a higher-bit quant. What it demonstrates is coherence at a footprint that fits a 16 GB Mac, and a working game written in a single pass.

Links

• Project website: mlx-optiq.com
• All OptiQ quants: mlx-optiq.com/models
• PyPI: pypi.org/project/mlx-optiq
• Base model: poolside/Laguna-S-2.1