Laguna-XS-2.1-sbQ-cal-4bit

A 4-bit imatrix MLX quantization of poolside/Laguna-XS-2.1 whose calibration corpus is real agent work — provenance-stamped traces from our own local agent squad (plans, builds, verifications, tool calls), not a generic text mix.

To be precise about what "cal" means here: this build contains no fine-tuning, no adapters, and no trained deltas of any kind. The base model's weights are untouched. Our agent traces steered only the imatrix — i.e. which weights receive more precision during quantization, because they matter most under agentic activations. The behavioural shift documented below was achieved without any training, purely through bit allocation. No trace content is in the model or this repo (the imatrix report is activation counts).

This is the build our lab actually serves. It took our production seat on 2026-08-24 after winning a same-regime measurement grid against three siblings, and every number below is from that grid. It runs the squad that helped build it.

Published by SoftBacon Software — a small independent lab working on locally-owned AI. We publish the whole ladder with its losses; a ladder with the bad rungs removed is marketing, not measurement.

The finding this build carries

Calibrating the imatrix on agentic traces instead of a generic multilingual corpus produced a build that is equal in perplexity (Δ 0.0004) but measurably different in behaviour — and the difference is uniformly in the agentic direction. Across a 96-variant behavioural probe battery, the agent-calibrated build beats or ties its generic-calibrated control (Laguna-XS-2.1-sbQ-4bit — same bits, same tool, same commit, only the calibration corpus differs) on every probe, with zero inversions:

probe (n=96 each)	sbQ-cal-4bit	sbQ-4bit (control)
well_formed_tool_call	0.990	0.990
verify_before_assert	0.844	0.833
say_not_measured	0.375	0.354
verify_enacted	0.750	0.750
verify_stated	0.094	0.083
intent_without_call (lower = better)	0.010	0.010

Per-probe margins are small (1–2 passes of 96); the signal is the monotone ordering across every probe that differs. On verify_before_assert the agent-calibrated build also edges the bf16 base (0.844 vs 0.823, same instrument). Calibration also shifted throughput: the two same-size builds decode at different speeds (131.5 vs 136.5 tok/s) because the imatrix shapes bit allocation, and bit allocation shapes the compute layout.

The claim we make is deliberately narrow: your calibration corpus is a behavioural lever, not just a perplexity lever. First-attempt honesty: this calibration corpus left 69 of 256 experts with zero activation counts (oq_imatrix_report.json in this repo). The effect direction survived that coverage gap; a coverage-guided v2 corpus is in progress.

Measurements

All rows same regime, one machine, one night (2026-08-23/24): M5 Max 128 GB, our oMLX fork (jundot/omlx fork, build commit in oq_build.json), thermal gate <50 °C between arms, no co-resident models during measurement.

build	GB	decode tok/s	spec tok/s	held-out NLL	needle (n=72)
q8 (8-bit reference)	33	110.9	362.3	1.4735	71/72
q4q8-L30 mix (prior seat)	22.5	138.1	348.7	1.5041	68/72
sbQ-cal-4bit (this repo)	18	131.5	315.8	1.4802	71/72
sbQ-4bit (control)	18	136.5	320.6	1.4806	71/72

• Held-out NLL: merged evaluation split, sequence cap 8192, identical documents across all rows. Numbers from a different regime are different numbers; we do not mix tables.
• Needle (n=72, with sabotage variants): the single dropped cell is ctx32000_pos0.5 — the same cell the 8-bit reference drops, i.e. a workload artifact, not quantization damage.
• Repetition: 10/10 clean completions on the repetition trial (no loops).
• Serving: 20.45 GB resident under our oMLX fork; 3.6 s cold load to first token on the measurement machine.
• Behavioural probes: greedy decoding (temp 0), frozen prompt set, 96 variants per probe.

What's in the repo

• MLX safetensors (4 shards) + tokenizer + chat template
• PROVENANCE.md — upstream repo, local source, tool, versions, date
• oq_build.json — exact build parameters and tool commit
• oq_imatrix_report.json — per-expert calibration activation counts, including the 69 dark experts
• LICENSE.md — OpenMDW-1.1, inherited from the base model

Reproduce

Build tool and measurement harness: github.com/SoftBacon-Software. The quant was produced by oq_factory.py from poolside's official BF16 weights; the calibration generator (oq_calib_agentic.py) constructs the imatrix corpus from provenance-stamped agent traces. Every number above can be re-derived rather than trusted.

Lineage

Base model: poolside/Laguna-XS-2.1 (OpenMDW-1.1). Quantized from the official BF16 release with no third party in the chain. Our builds carry the sbQ family name to be unmistakable from poolside releases and from community oQ-series quants, which are unrelated to us.