Soulfate24/Paretrix_Quantization_Suite

Paretrix Quantization Suite (v2.6.0)

Paretrix is an empirical, activation-aware quantization suite for GGUF language models and speculative decoding modules.

Pareto + Matrix → Paretrix. Each tier is priced so that no better trade exists at that compression. The suite measures real activation sensitivity (ΔKLD/MiB) per tensor class through llama-imatrix, learns rate tables from cross-architecture campaigns (THE MATRIX), and allocates bitwidths under exact budget targets: flat recipes where uniformity wins, rate-calibrated knapsack where heterogeneity pays.

Standard uniform quantization applies one bitwidth family across the whole network. Paretrix reads the model's own activation field (which classes are expensive to cut, which are nearly free, and where depth matters), then spends the budget where measurements show the greatest return.

---

🪜 The Mod-3 Ladder

Each tier lands on a multiple of 3 percent of the unquantized BF16 source, so the shipping name states its weight band (model-Paretrix-<Tier>-<XX>pc.gguf, ±1.5 pp). Tiers run from largest to smallest.

Tier	Ratio	Allocation Strategy	Primary Purpose
Fidelity	48%	Q6_K base + Q8_0 pockets biased to the late layer third (G4)	High-fidelity archival tier.
Precision	42%	Flat Q6_K + full-attention band Q8_0; the readout lever is arch-scoped (L12)	Coding, mathematics, high-entropy reasoning.
Quality	36%	Flat Q5_K_M + imatrix (L6); self re-shuffle where the twin is a stock copy	General production tier.
Compact	33%	Exact-budget DP, or flat Q4_K_M + band (Q6_K) + recurrent floors	Daily driver for 2B–4B architectures.
Mini	30%	Exact-budget DP, or L4 flat on GDN hybrids (IQ4_XS + recurrent floors + full-attention band)	Balanced entry tier for ≥ 9B models on 8 GB VRAM.
Nano	27%	Exact-budget DP (IQ3_S–IQ4_XS base, tied readout Q5_K when vocab ≥ 15% of mass)	Compact edge deployment below stock IQ4_XS.
Pico	24%	Exact-budget DP (IQ3_S–IQ4_XS base, scarce-attention band protected)	Edge budget; default on state-rich ≥3B trunks, explicit elsewhere (PARETRIX_INCLUDE_PICO=1).
Femto	21%	Exact-budget DP, deep-squeeze product (IQ3_XXS base, measured-rate squeezes)	Deepest service rung; campaign product, default on state-rich ≥3B trunks.

Stock twins. On topologies where a flat recipe's levers have no effect under the regex rules, the Paretrix tier is byte-identical to its stock twin (detected automatically). The suite then queues one self re-shuffle campaign anchored on the twin. Measured outcome: the re-shuffle wins below the top band and ties at it (R19).

---

🧮 THE MATRIX — measured rate knowledge

Three pricing levels, each overlaying the next, class by class:

1. Model cache (model-paretrix.json): this model's own measured tables, anchored on the operating point nearest the target (R20).
2. Arch line (arch-paretrix/<arch>-paretrix.json): shared per-architecture knowledge, including banded mid/deep sell and buy tables, structural facts, build corrections, and witnesses. Size-normalized (rate × line_reference / model_size), so checkpoints of different scales share one line.
3. Universal prior (arch-paretrix/generic-paretrix.json): the distilled six-family prior, also banded and size-aware.

The fold (Paretrix.py matrix) compiles every discoverable model cache into the arch lines:

• Sells price at the measured range maximum, which stays conservative for every cut.
• Buys price at the minimum of the range lows, which stays conservative for every gain.
• Dead buy classes price at zero (R13), so the DP keeps the measured zero instead of falling back to the MSE proxy.

---

⚔️ Campaign engine (RCO)

One command per rung: anchor → notch probes → measured rates → exact-budget DP → build → duel at the verdict regime → adopt.

python Paretrix.py campaign --model model-BF16.gguf --imatrix imatrix.gguf \
    --anchor-from-artifact model-Q5_K_M-imx.gguf --target-mib 1062 --run
python Paretrix.py duel --model model-BF16.gguf --tier quality \
    --candidate model-Paretrix-Candidate-36pc.gguf --adopt

The DP solves the multiple-choice knapsack to the exact MiB, with no constraint hyperparameter. Squeezes below the floors are priced from the measured notch rates; raises above the anchor are priced from the measured buy table, with dead buys clamped to zero. The duel binds at ctx 4096×8 and adopts on a KLD gain at or above 1×MASD (floor 0.001), the same gate the Pareto column reads.

---

📊 Benchmarks

All sweeps run at the verdict regime ctx 4096×8 (32,768-token span) on the corpus wiki.test.raw with Flash-Attention. The KL reference is the BF16 logits of the same model. Compare KLD only within a single regime.

Models run from largest to smallest by BF16 size, and each table runs from largest to smallest MiB.

Pareto column: ★ = on this sweep's (MiB, KLD) frontier · < X = this row is dominated by X (X is lighter and better) · ≈ X = X is lighter, with KLD inside the tie zone max(MASD, 0.001) · ≡ = byte-identical twin.

---

1. ornith-ai/Ornith-1.5-9B — qwen35 GDN hybrid · untied 248k vocab

BF16 17,091 MiB · target: 8+ GB VRAM (split offload) · PPL reference 8.9219 · DFlash draft: ×1.34–×2.14 net speedup.

Model	MiB	PPL	ΔPPL	KLD	RMS Δp	top-p	Pareto
Q8_0 (stock)	9086	8.9544	+0.0325	0.0118	2.67%	97.7%	★
Fidelity-48pc	8232	8.9700	+0.0481	0.0167	3.13%	97.0%	≈ Precision-42pc (−0.0011 · −854 MiB)
Precision-42pc	7378	8.9365	+0.0146	0.0156	3.24%	96.6%	★
Q6_K-imx (stock)	7018	8.7000†	−0.2219	0.0251	3.83%	95.9%	★
Quality-36pc	6234	8.5716†	−0.3503	0.0449	4.88%	93.5%	★
Q5_K_M-imx (stock)	6168	8.2043†	−0.7176	0.1013	7.29%	90.2%	< Compact-33pc
Compact-33pc	5787	8.4619†	−0.4600	0.0610	6.07%	92.0%	★
Mini-30pc ⭐	5368	8.6772†	−0.2447	0.0679	6.42%	91.1%	★
IQ4_XS-imx (stock)	4956	9.2939	+0.3720	0.0794	7.08%	90.8%	★
Nano-27pc	4586	8.6546†	−0.2673	0.1223	8.73%	86.5%	★
IQ3_M-imx (stock)	4211	9.1809	+0.2590	0.1807	11.04%	84.5%	< Pico-24pc
Pico-24pc	4070	8.5303†	−0.3916	0.1540	10.08%	84.7%	★
Femto-21pc	3601	8.3140†	−0.6079	0.2524	12.37%	80.6%	★

Two strict dominations, the program's largest: Compact (−0.0403 KLD, −381 MiB) and Pico (−0.0268 KLD, −141 MiB). Fidelity ties Precision within the tie zone while weighing 854 MiB more. At the top band, the extra Q8_0 pockets buy nothing measurable (L16).

---

2. TokenRhythm/NeoHorse-1-4B — qwen3.5 fine-tune · GDN hybrid · tied

BF16 8,034 MiB · target: 4+ GB VRAM · PPL reference 9.1268 · L19 fine-tune field (7/7 dead buys at mid anchors).

Model	MiB	PPL	ΔPPL	KLD	RMS Δp	top-p	Pareto
Q8_0 (stock)	4275	9.1273	+0.0005	0.0057	1.82%	98.2%	★
Fidelity-48pc	3866	9.1807	+0.0539	0.0083	2.45%	97.5%	≈ Precision-42pc (+0.0002 · −494 MiB)
Precision-42pc	3372	9.1828	+0.0560	0.0084	2.82%	97.0%	≈ Q6_K-imx (+0.0009 · −68 MiB)
Q6_K-imx (stock)	3304	9.1991	+0.0724	0.0093	2.90%	96.7%	★
Quality-36pc	2966	9.2173	+0.0905	0.0204	3.71%	95.1%	★
Q5_K_M-imx (stock)	2933	9.3403	+0.2135	0.0270	4.32%	94.3%	★
Compact-33pc	2768	9.1505	+0.0237	0.0300	4.71%	93.1%	★
Mini-30pc	2416	9.5442	+0.4174	0.0537	6.02%	90.9%	★
IQ4_XS-imx (stock)	2398	9.5261	+0.3994	0.0598	6.37%	90.5%	★
Nano-27pc ⭐	2251	9.2769	+0.1501	0.0683	6.85%	89.7%	★
IQ3_M-imx (stock)	2063	9.9031	+0.7763	0.1480	10.62%	84.8%	★
Pico-24pc	1935	10.1509	+1.0242	0.1570	10.92%	84.6%	★
Femto-21pc	1696	10.0878	+0.9610	0.2534	13.98%	79.3%	★

Precision matches Fidelity within the tie zone at 494 MiB less. The fine-tune field flattens sells (×0.3) and leaves most mid-band buys dead (L19), so campaign value concentrates in deep sheds and self re-shuffles.

---

3. Nanbeige/Nanbeige4.2-3B — looped dense · untied 166k vocab

BF16 7,957 MiB · target: 6+ GB VRAM · PPL reference 29.8955 · readout pair 24.5% of mass · R3 loop-neutral depth prior active · DSpark draft: ×1.87/×2.13 net speedup.

Model	MiB	PPL	ΔPPL	KLD	RMS Δp	top-p	Pareto
Q8_0 (stock)	4229	31.5657	+1.6702	0.0144	3.01%	95.5%	★
Fidelity-48pc	3821	31.8994	+2.0039	0.0266	3.97%	93.8%	★
Precision-42pc	3266	32.0971	+2.2016	0.0467	4.93%	91.9%	≡ Q6_K-imx
Q6_K-imx (stock)	3266	32.0971	+2.2016	0.0467	4.93%	91.9%	≡ Precision-42pc
Q5_K_M-imx (stock)	2849	32.0553	+2.1598	0.0872	6.58%	88.2%	< Quality-36pc
Quality-36pc ⭐	2846	31.7395	+1.8440	0.0714	5.86%	89.2%	★
Compact-33pc	2629	31.7929	+1.8974	0.1115	7.53%	86.5%	★
Mini-30pc	2388	30.7090	+0.8135	0.1841	9.43%	82.8%	★
IQ4_XS-imx (stock)	2268	32.7700	+2.8745	0.2093	10.24%	81.7%	★
Nano-27pc	2152	31.3016	+1.4061	0.2287	10.51%	80.4%	★
IQ3_M-imx (stock)	1985	33.2114	+3.3158	0.4948	15.12%	71.2%	< Pico-24pc
Pico-24pc	1906	30.0424†	+0.1469	0.4023	14.04%	73.2%	★
Femto-21pc	1675	36.4632	+6.5677	0.5857	16.35%	68.3%	★

Two strict dominations: Quality (−0.0158 KLD, −3 MiB) and Pico (−0.0925 KLD, −79 MiB). Precision is byte-identical to Q6_K-imx (detected by R19).

---

4. XHToken/Spark-X2.5-4B — dense SWA hybrid 3:1 · tied 131k vocab

BF16 7,849 MiB · target: 6+ GB VRAM · PPL reference 20.1134 · the verdict binds at native context (L18).

Model	MiB	PPL	ΔPPL	KLD	RMS Δp	top-p	Pareto
Q8_0 (stock)	4172	19.9740	−0.1394	0.0047	1.60%	97.3%	★
Fidelity-48pc	3772	20.0099	−0.1035	0.0115	2.49%	95.3%	★
Precision-42pc	3359	19.7049	−0.4085	0.0128	2.66%	95.0%	★
Q6_K-imx (stock)	3223	19.6832	−0.4303	0.0173	3.02%	93.8%	★
Q5_K_M-imx (stock)	2840	20.8885	+0.7751	0.0571	5.57%	89.4%	< Quality-36pc
Quality-36pc ⭐	2838	20.0193	−0.0941	0.0437	4.70%	90.4%	★
Compact-33pc	2587	20.6036	+0.4902	0.0976	7.16%	86.1%	★
Mini-30pc	2343	21.3848	+1.2714	0.1203	8.06%	84.6%	★
IQ4_XS-imx (stock)	2266	23.2522	+3.1388	0.1678	9.45%	81.8%	★
Nano-27pc	2123	22.3853	+2.2719	0.2498	11.35%	78.4%	★
Pico-24pc	1982	19.4291†	−0.6843	0.2812	12.30%	76.9%	★
IQ3_M-imx (stock)	1949	21.6523	+1.5389	0.3635	14.23%	73.7%	★
Femto-21pc	1664	26.5109	+6.3975	0.5799	17.22%	67.7%	★

Quality dominates Q5_K_M-imx (−0.0134 KLD, −2 MiB). The spark2_5 corrections (readout lever L12, F32 gate pin R9, ffn_down Q4_K lever) apply to every build; the scarce 9-layer full-attention band is the family's richest buy.

---

5. webAI-Official/TwIL-LM3-Pro — granite dense · untied 128k vocab

BF16 6,984 MiB · target: 6+ GB VRAM · PPL reference 11.4980 · R19 reference case.

Model	MiB	PPL	ΔPPL	KLD	RMS Δp	top-p	Pareto
Q8_0 (stock)	3712	11.5097	+0.0117	0.0020	1.33%	97.9%	≈ Fidelity-48pc (+0.0007 · −357 MiB)
Fidelity-48pc	3355	11.5074	+0.0094	0.0027	1.38%	97.4%	★
Precision-42pc	2867	11.5042	+0.0062	0.0049	1.99%	96.6%	≡ Q6_K-imx
Q6_K-imx (stock)	2867	11.5042	+0.0062	0.0049	1.99%	96.6%	≡ Precision-42pc
Quality-36pc	2514	11.6187	+0.1207	0.0104	2.74%	95.1%	★
Q5_K_M-imx (stock)	2493	11.6684	+0.1704	0.0131	3.33%	94.5%	★
Compact-33pc	2306	11.6594	+0.1614	0.0218	3.90%	93.2%	★
Mini-30pc	2096	11.6689	+0.1709	0.0300	4.59%	91.6%	★
IQ4_XS-imx (stock)	1937	11.8072	+0.3092	0.0350	4.92%	91.1%	★
Nano-27pc ⭐	1885	11.9953	+0.4973	0.0479	5.64%	89.4%	★
IQ3_M-imx (stock)	1653	12.4715	+0.9735	0.1052	8.63%	84.7%	★

R19 case: Quality-36pc and Precision-42pc started byte-identical to their stock twins. The self re-shuffle on Quality won −20.7% (0.0131 → 0.0104, 2.7× the duel gate). On Precision it tied (+0.0007, inside the gate, so the incumbent stands). The re-shuffle wins or ties by construction.

---

6. LiquidAI/LFM2.5-2.6B — shortconv mixer · tied readout

BF16 5,153 MiB · target: 4+ GB VRAM · PPL reference 51.4190 · DSpark draft: ×1.10 net speedup.

Model	MiB	PPL	ΔPPL	KLD	RMS Δp	top-p	Pareto
Q8_0 (stock)	2742	51.8867	+0.4677	0.0032	1.34%	97.4%	★
Fidelity-48pc	2481	52.2324	+0.8134	0.0076	2.09%	96.2%	★
Precision-42pc	2138	52.8312	+1.4122	0.0105	2.44%	95.4%	★
Q6_K-imx (stock)	2119	52.7076	+1.2886	0.0131	2.66%	94.7%	★
Q5_K_M-imx (stock)	1850	51.5220	+0.1030	0.0380	4.39%	91.0%	< Quality-36pc
Quality-36pc ⭐	1848	52.6558	+1.2368	0.0309	4.13%	91.6%	★
Compact-33pc	1712	49.4337	−1.9853	0.0696	6.13%	88.0%	★
Mini-30pc	1560	46.9409†	−4.4780	0.0995	7.45%	85.9%	★
IQ4_XS-imx (stock)	1447	52.6602	+1.2412	0.1652	9.56%	81.6%	< Nano-27pc
Nano-27pc	1405	52.7926	+1.3736	0.1536	8.91%	81.8%	★
Pico-24pc	1273	55.9255	+4.5065	0.3597	13.43%	74.6%	★
IQ3_M-imx (stock)	1225	60.9679	+9.5490	0.4254	14.31%	72.6%	★
Femto-21pc	1096	71.9524	+20.5334	0.5535	16.29%	69.0%	★

Quality dominates Q5_K_M-imx (−0.0071 KLD, −2 MiB), and Nano dominates IQ4_XS-imx (−0.0116 KLD, −42 MiB). † PPL below the BF16 base indicates entropy collapse, not a quality gain.

---

7. openbmb/MiniCPM5-2B — dense llama-arch · untied 130k readout

BF16 4,806 MiB · target: 4+ GB VRAM · PPL reference 11.9212 · DSpark draft: acceptance 0.4419, ×1.89 net speedup.

Model	MiB	PPL	ΔPPL	KLD	RMS Δp	top-p	Pareto
Q8_0 (stock)	2556	11.9393	+0.0181	0.0015	0.99%	97.8%	★
Fidelity-48pc	2309	11.9598	+0.0386	0.0033	1.47%	96.8%	★
Q6_K-imx (stock)	1974	11.9724	+0.0512	0.0059	1.91%	95.9%	≈ Precision-42pc (−0.0002 · −1 MiB)
Precision-42pc	1973	11.9678	+0.0466	0.0057	1.89%	96.0%	★
Q5_K_M-imx (stock)	1724	12.0905	+0.1693	0.0195	3.58%	92.8%	< Quality-36pc
Quality-36pc	1723	12.1180	+0.1968	0.0175	3.43%	93.3%	★
Compact-33pc ⭐	1593	12.2219	+0.3007	0.0427	5.12%	89.2%	★
Mini-30pc	1446	12.3731	+0.4519	0.0672	6.53%	86.7%	★
IQ4_XS-imx (stock)	1358	12.3977	+0.4765	0.0722	6.69%	86.9%	★
Nano-27pc	1306	12.6530	+0.7318	0.0848	7.26%	85.7%	★
IQ3_M-imx (stock)	1170	13.9101	+1.9889	0.2027	11.74%	78.6%	< Pico-24pc
Pico-24pc	1154	13.9155	+1.9943	0.1719	10.52%	79.3%	★
Femto-21pc	1016	15.2794	+3.3582	0.2918	14.25%	73.9%	★

Two strict dominations: Quality (−0.0020 KLD, −1 MiB) and Pico (−0.0308 KLD, −15 MiB). All eight rungs held through the L17 self re-shuffle wave.

---

8. jinaai/ReaderLM-v2 — dense qwen2 · tied 152k vocab · HTML/Markdown extractor

BF16 2,950 MiB · target: 2+ GB VRAM · PPL reference 12.3852 · FFN holds 75% of the mass.

Model	MiB	PPL	ΔPPL	KLD	RMS Δp	top-p	Pareto
Q8_0 (stock)	1570	12.4136	+0.0284	0.0016	0.99%	97.9%	★
Fidelity-48pc	1422	12.4161	+0.0309	0.0031	1.39%	96.9%	★
Precision-42pc	1214	12.3712†	−0.0140	0.0055	1.84%	95.9%	≡ Q6_K-imx
Q6_K-imx (stock)	1214	12.3712†	−0.0140	0.0055	1.84%	95.9%	≡ Precision-42pc
Quality-36pc	1073	12.4244	+0.0392	0.0165	3.25%	93.1%	≡ Q5_K_M-imx
Q5_K_M-imx (stock)	1073	12.4244	+0.0392	0.0165	3.25%	93.1%	≡ Quality-36pc
Compact-33pc ⭐	979	12.4617	+0.0765	0.0309	4.35%	90.9%	★
Mini-30pc	890	12.6359	+0.2507	0.0553	5.95%	87.7%	≈ IQ4_XS-imx (+0.0027 · −36 MiB)
IQ4_XS-imx (stock)	854	12.5393	+0.1541	0.0580	5.96%	87.6%	★
Nano-27pc	802	12.6418	+0.2566	0.0887	7.39%	84.9%	★
IQ3_M-imx (stock)	741	13.3774	+0.9922	0.1706	10.37%	79.6%	★

Pico and Femto are omitted: the constructibility gate applies (tied vocab sits at 15% of the mass, the R6-ter boundary). Compact, Mini and Nano are campaign products. Quality and Precision are byte-identical to their stock twins (detected by R19). L2 validated: the Q4_K ffn_down floor restored monotonicity across the 75% FFN mass.

---

9. SeerRay-Lab/Xiaomi-OCR-0 — qwen35 GDN hybrid · tied 248k vocab

BF16 1,446 MiB · target: 2+ GB VRAM · PPL reference 28.5189 · vision via the mmproj module.

Model	MiB	PPL	ΔPPL	KLD	RMS Δp	top-p	Pareto
Q8_0 (stock)	774	28.5404	+0.0215	0.0010	0.77%	98.1%	★
Fidelity-48pc	704	28.5178	−0.0011	0.0022	1.12%	97.0%	★
Precision-42pc	611	28.5667	+0.0477	0.0050	1.68%	95.4%	≈ Q6_K-imx (+0.0003 · −10 MiB)
Q6_K-imx (stock)	601	28.6110	+0.0920	0.0052	1.74%	95.4%	★
Q5_K_M-imx (stock)	551	28.9326	+0.4136	0.0111	2.51%	93.7%	★
Quality-36pc	538	28.9461	+0.4271	0.0131	2.72%	92.8%	★
Compact-33pc ⭐	487	29.3475	+0.8285	0.0289	3.70%	90.1%	★
IQ4_XS-imx (stock)	481	29.6333	+1.1144	0.0425	4.81%	87.7%	★
Mini-30pc	444	29.8250	+1.3060	0.0521	5.18%	86.7%	★
IQ3_M-imx (stock)	433	32.1805	+3.6615	0.1050	8.26%	81.0%	★

Pico and Femto are omitted: the constructibility gate applies (tied vocab is 34% of the mass, and the Q6_K floor alone reaches 51% of the Nano target). Mini-30 is the family floor. Precision ≈ Q6_K-imx within the tie zone (C2). The pipeline skips this rung on small qwen35 by default; set PARETRIX_PRECISION_TWIN=0 to force it.

---

📉 Sub-Nano compendium — the Nano → Pico step

Families run from largest to smallest.

Family	Architecture	Nano KLD	Pico KLD	Step	Pico verdict
Ornith-1.5-9B	GDN hybrid, untied	0.1223	0.1540	+26%	✅ Fully usable (beats IQ3_M 0.1807)
NeoHorse-1-4B	GDN fine-tune, tied	0.0683	0.1570	+130%	◐ Context variant (IQ3_M reaches 0.1480 at +128 MiB)
Nanbeige4.2-3B	Looped dense, untied	0.2287	0.4023	+76%	✗ Degraded (Nano already past the 0.20 edge)
Spark-X2.5-4B	SWA hybrid, tied	0.2498	0.2812	+13%	✗ Degraded (floor = Mini)
TwIL-LM3-Pro	Dense, untied	0.0479	—	—	⊘ Ladder stops at Nano
LFM2.5-2.6B	Shortconv mixer, tied	0.1536	0.3597	+134%	✗ Pico exceeds 0.20 (floor = Mini)
MiniCPM5-2B	Dense, untied	0.0848	0.1719	+103%	✅ Usable (beats IQ3_M 0.2027)
ReaderLM-v2	Dense, tied	0.0887	—	—	⊘ Constructibility gate (floor = Nano)
Xiaomi-OCR-0	GDN hybrid, tied	—	—	—	⊘ Constructibility gate (floor = Mini)

---

Summary — 9 strict Pareto dominations

Products that are both lighter and better than their stock twin at the verdict regime. Families run from largest to smallest; tiers run from largest to smallest within each family.

Family	Product	KLD	Stock twin	Stock KLD	ΔKLD	ΔMiB
Ornith-1.5-9B	Compact-33pc	0.0610	Q5_K_M-imx	0.1013	−0.0403	−381
Ornith-1.5-9B	Pico-24pc	0.1540	IQ3_M-imx	0.1807	−0.0268	−141
Nanbeige4.2-3B	Quality-36pc	0.0714	Q5_K_M-imx	0.0872	−0.0158	−3
Nanbeige4.2-3B	Pico-24pc	0.4023	IQ3_M-imx	0.4948	−0.0925	−79
Spark-X2.5-4B	Quality-36pc	0.0437	Q5_K_M-imx	0.0571	−0.0134	−2
LFM2.5-2.6B	Quality-36pc	0.0309	Q5_K_M-imx	0.0380	−0.0071	−2
LFM2.5-2.6B	Nano-27pc	0.1536	IQ4_XS-imx	0.1652	−0.0116	−42
MiniCPM5-2B	Quality-36pc	0.0175	Q5_K_M-imx	0.0195	−0.0020	−1
MiniCPM5-2B	Pico-24pc	0.1719	IQ3_M-imx	0.2027	−0.0308	−15

Plus four byte-identical stock twins detected (R19): Nanbeige Precision ≡ Q6_K-imx · TwIL Precision ≡ Q6_K-imx · ReaderLM Precision ≡ Q6_K-imx · ReaderLM Quality ≡ Q5_K_M-imx.

---

⚡ Speculative decoding: MTP · DSpark · DFlash

Paretrix quantizes speculative modules imatrix-free (Paretrix-modules.py). A draft's input is the parent's hidden state, the MTP head rides the trunk's activations, and the vision tower consumes raw pixels, so none of them exposes a corpus entry point. Speculative verification bounds draft damage to the acceptance rate and leaves output fidelity intact (L11).

Module	Role	Palette	Naming
MTP	Multi-token prediction fusion head (extracted by 01)	Role floors + K-quants	mtp-Paretrix-<Profile>.gguf
DSpark	Block-diffusion speculative draft (Markov + confidence heads)	Role floors + K-quants	dspark-Paretrix-<Profile>.gguf
DFlash	Third-party block-diffusion draft	Role floors + K-quants	dflash-Paretrix-<Profile>.gguf
MMProj	CLIP vision projector	Profile grid, critical tensors pinned F32	mmproj-Paretrix-<Profile>.gguf

Measured acceptance invariance (L11). Draft quantization moves acceptance by no more than ±0.03: nanbeige 0.3988 → 0.4092 (Δ +0.0104) · MiniCPM5-2B 0.4419 (ledger 0.4464) · lfm2 +0.0055 · Ornith DFlash 0.2631 → 0.2684 at matched ngl. The quantized draft also runs faster: net speedups of ×1.87/×2.13 (nanbeige), ×1.89 (MiniCPM5-2B), and ×1.34–×2.14 (Ornith DFlash, 106.9 t/s on a 9B model at 8 GB VRAM). Acceptance is a property of the (draft, target) pair and the offload regime (L20), so compare acceptance only at matched --ngl.

# Acceptance battery (llama-server A/B, temperature 0, trained block auto-read)
python 13_draft-acceptance-sweep.py <working-folder> --chat --tokenizer-dir <hf-folder>

# Fuse model + modules into one deployment GGUF
python 14_gguf-module-fusion.py fused.gguf model-BF16.gguf mmproj-BF16.gguf

---

⚡ Quick start

# One command: convert → imatrix → full Paretrix ladder [+ classic line]
python 00_Paretrix-pipeline.py ./MyModel [--all]

# Or step by step
python 01_SAFETENSORS-to-BF16-GGUF.py ./MyModel  # → Paretrix/model-BF16.gguf
python 02_BF16-GGUF-to-Q8-imatrix.py             # → model-Q8_0.gguf + imatrix.gguf
python 03_BF16-GGUF-to-Paretrix.py [--all]       # → full ladder + duels

# Campaign on a specific rung (anchor → probes → DP → build → duel → adopt)
python Paretrix.py campaign --model model-BF16.gguf --imatrix imatrix.gguf \
    --anchor-from-artifact model-Q5_K_M-imx.gguf --target-mib 1062 --run

# Sweep every product at the verdict regime
python 11_perplexity-test.py

# Fold the model caches into the shared arch lines
python Paretrix.py matrix

Working-root convention. A model folder holds its HF repo (safetensors + sidecars) plus a Paretrix/ subfolder carrying every product: BF16/Q8/imatrix GGUFs, champions, modules, and model-paretrix.json.

<model folder>/
├─ model-00001-of-00002.safetensors · config.json · tokenizer.json …
└─ Paretrix/
   ├─ model-BF16.gguf · model-Q8_0.gguf · imatrix.gguf
   ├─ model-Paretrix-<Tier>-<XX>pc.gguf · mtp/mmproj/dspark products
   └─ model-paretrix.json

Full operational detail: DOCUMENTATION.md.

---

📜 Citation & Credits

• Paretrix Quantization Suite: the Pareto-Matrix rate-exchange framework, with measured notch rates, exact-budget DP allocation, and a cross-architecture knowledge base (THE MATRIX).
• llama.cpp by Georgi Gerganov & ggml contributors: GGUF/GGML runtime, llama-imatrix, llama-quantize, llama-perplexity, and llama-server.
• GSQ / RCO (IST-DASLab): GSQ: Gumbel-STE quantization (arXiv:2604.18556) · RCO: relaxed knapsack allocation (arXiv:2605.00649). The exact-budget DP and Gumbel-STE manifold search build on these lines, adapted to measured rates.
• wepiqx ASHQ1: priority-queue knapsack formulation, tied-group activation hashing, and MSE scheduling.
• Empero AI: GDN state preservation (ssm_alpha/ssm_beta at Q8_0).