AtomicChat/Laguna-XS-2.1-MLX-8bit

Base model: poolside/Laguna-XS-2.1

Laguna XS 2.1, quantized to MLX (8-bit) by Atomic Chat for Apple Silicon. Built straight from poolside's original weights. Runs fully offline on your Mac.

Highlights

• 33B total / 3B active Mixture-of-Experts for agentic coding and long-horizon work on a local machine.
• Mixed attention layout: 40 layers, 10 global + 30 sliding-window (3:1 ratio), sigmoid gating with per-layer rotary scales.
• 256 experts + 1 shared expert, sliding window of 512 tokens.
• 262,144-token context.
• Native interleaved reasoning, enable or disable per request.
• Upgraded from Laguna XS.2: +5.4% on SWE-bench Multilingual and stronger terminal-style performance.

These are MLX builds for Apple Silicon (M-series), quantized from the original weights, not a repack. Laguna's architecture runs on mlx-vlm (0.6.3+) as a text model; stock mlx-lm does not yet include it.

Model Overview

Property	Value
Base model	poolside/Laguna-XS-2.1
Total parameters	33B (3B active per token)
Architecture	Laguna MoE, mixed sliding-window/global attention
Experts	256 + 1 shared
Layers	40 (10 global, 30 sliding-window)
Sliding window	512 tokens
Context length	262,144
Optimizer	Muon
This repo	MLX quants (3-8 bit) for Apple Silicon, built from the original weights with mlx-vlm.

Scores are poolside's published results for the full-precision base poolside/Laguna-XS-2.1. The MLX quants run the same model locally; lower bit-widths trade a little accuracy for size and speed.

This quant

This repo is the 8-bit MLX build (~33 GB). The full ladder (5/6/8-bit) lives in the Laguna XS 2.1 collection.

Get started

• Atomic Chat: open the app, search AtomicChat/Laguna-XS-2.1-MLX-8bit, pick a quant, hit Use this model.
• mlx-vlm (generate):

  pip install -U mlx-vlm
  python -m mlx_vlm generate --model AtomicChat/Laguna-XS-2.1-MLX-8bit-8bit \
      --prompt "Write a Python retry wrapper with exponential backoff." \
      --max-tokens 512 --temperature 1.0
  

• mlx-vlm (OpenAI-compatible server):

  python -m mlx_vlm server --model AtomicChat/Laguna-XS-2.1-MLX-8bit-8bit --host 0.0.0.0 --port 8080
  # POST http://localhost:8080/v1/chat/completions  with  "model": "6bit"
  

Reasoning is native and on by default. Start the server with --enable-thinking (optionally --thinking-budget N) to keep it; omit the flag for direct, non-reasoning replies.

Best practices

Parameter	Value
temperature	1.0
top_k	20
top_p	1.0

poolside's benchmark settings. For agentic coding, keep reasoning enabled and preserve prior thinking blocks across turns.

How these were made

1. Download poolside/Laguna-XS-2.1 (original BF16 weights).
2. Quantize each rung with python -m mlx_vlm convert --hf-path poolside/Laguna-XS-2.1 -q --q-bits <N> --q-group-size 64.

License

Released by poolside under the OpenMDW-1.1 license, which permits free use, modification and redistribution with attribution. MLX conversion by Atomic Chat. This is an unofficial community quantization and is not endorsed by poolside; the original LICENSE.md and notices of origin are retained in each quant folder.