| --- |
| license: openmdw-1.1 |
| base_model: poolside/Laguna-XS-2.1 |
| base_model_relation: quantized |
| library_name: mlx |
| pipeline_tag: text-generation |
| tags: |
| - mlx |
| - laguna |
| - quantized |
| - 4-bit |
| - imatrix |
| - sbq |
| - apple-silicon |
| --- |
| |
| # Laguna-XS-2.1-sbQ-4bit |
|
|
| A 4-bit imatrix MLX quantization of [poolside/Laguna-XS-2.1](https://hf.co/poolside/Laguna-XS-2.1), |
| calibrated on a generic code + multilingual corpus. |
|
|
| **This repo exists as the control.** It is the exact same build — same bits, |
| same tool, same commit — as |
| [Laguna-XS-2.1-sbQ-cal-4bit](https://hf.co/SoftBacon-Software/Laguna-XS-2.1-sbQ-cal-4bit), |
| except the imatrix calibration corpus: generic text here, real agent traces |
| there. The comparison between the two is the finding — calibration shifts |
| behaviour and speed at equal perplexity — and a comparison you cannot |
| reproduce is a story, so both sides are published. Read the sbQ-cal card for |
| the full grid and the probe battery; this card carries this build's own |
| numbers. |
|
|
| Published by [SoftBacon Software](https://hf.co/SoftBacon-Software). We |
| publish the whole ladder, losing rungs included. On behavioural probes this |
| build ties or loses to its agent-calibrated sibling on every probe (by 1–2 |
| passes of 96, zero inversions) — that is precisely why it is here. |
|
|
| ## Measurements (same regime as the sbQ-cal card, 2026-08-23/24) |
|
|
| | | this build | sbQ-cal-4bit | |
| |---|---|---| |
| | size | 18 GB | 18 GB | |
| | decode tok/s | **136.5** | 131.5 | |
| | speculative tok/s | **320.6** | 315.8 | |
| | held-out NLL (merged split, cap 8192) | 1.4806 | **1.4802** | |
| | needle n=72 | 71/72 | 71/72 | |
| | verify_before_assert (n=96) | 0.833 | **0.844** | |
| | say_not_measured (n=96) | 0.354 | **0.375** | |
|
|
| It is the slightly **faster** of the pair — calibration shapes bit |
| allocation, and bit allocation shapes the compute layout — so if you want |
| raw decode speed over agentic-behaviour margins, this is the rung to take. |
| The single dropped needle cell (`ctx32000_pos0.5`) is dropped by the 8-bit |
| reference too: workload artifact, not quant damage. |
|
|
| Full expert-activation report from calibration: `oq_imatrix_report.json` |
| (0 dark experts of 256 for this corpus). |
|
|
| ## What's in the repo |
|
|
| MLX safetensors + tokenizer + chat template, `PROVENANCE.md`, |
| `oq_build.json`, `oq_imatrix_report.json`, and `LICENSE.md` (OpenMDW-1.1, |
| inherited from the base model). |
|
|
| ## Reproduce |
|
|
| Build tool and measurement harness: [github.com/SoftBacon-Software](https://github.com/SoftBacon-Software). |
| Built from poolside's official BF16 with no third party in the chain. The |
| `sbQ` family name marks SoftBacon builds — unrelated to poolside releases |
| and to community `oQ`-series quants. |
|
|
