Sage-A1.5 VAL 1M I-Matrix GGUF

Independently computed using dedicated cloud compute clusters, funded 100% out-of-pocket. If you consider these releases to be of high quality and they help you in your day-to-day workflow, consider supporting on Ko-fi. Every contribution directly sustains future cloud compute hours and helps support my university studies.

Technical Overview

Handcrafted Importance Matrix (I-Matrix) computed for InternScience/Sage-A1.5 to power high-fidelity downstream quantization (IQ4_XS, IQ3_M, IQ3_S, IQ2_XS, etc.).

Calculated using official llama.cpp b11512 (commit a11f57ba93797579a5d1855ee216a31f10242676) with full layer offloading (-ngl 99).

Parameter	Specification
Model Repository	`InternScience/Sage-A1.5`
Model Revision	`3e6b02c5e6afbc481fd4556baca8e6009f26d696`
Model Architecture	`qwen35moe` (40 layers, 256 experts per layer, top-k=8, intermediate=512)
Target Tokens	1,000,000 tokens (489 chunks x 2048 context = 1,001,472 positions)
Execution Flags	`--process-output --no-ppl --no-escape -c 2048 -b 2048 -ub 2048 --chunks 489 -ngl 99`

Macro Component Distribution (1,001,472 Positions)

The calibration dataset was assembled with a deterministic 35/25/25/15 ratio spanning 11 natural languages and 10 programming languages, fully verified against the native tokenizer:

Component	Target Share	Measured Share	Effective Tokens	Sub-categories / Scope
Code	35.0%	34.99%	350,419	10 Programming Languages (Python, JS, TS, Java, C++, C, Go, Rust, SQL, Shell)
Reasoning	25.0%	24.83%	248,641	11 Natural Languages (Mathematical, logical, and step-by-step reasoning)
Prose / Text	25.0%	25.06%	251,005	11 Natural Languages (High-density prose, encyclopedic text, technical documentation)
Tools / Agentic	15.0%	15.12%	151,407	11 Natural Languages (Multi-turn tool transcripts, function calling, API schemas)
Total Positions	100.0%	100.0%	1,001,472	489 chunks of 2048 tokens

Detailed Token Breakdown by Language & Category

1. Code Breakdown by Programming Language (350,419 Tokens)

Programming Language	Effective Tokens	Category Share	Corpus Share
Python	92,660	26.44%	9.25%
JavaScript	57,988	16.55%	5.79%
C++	35,351	10.09%	3.53%
TypeScript	35,216	10.05%	3.52%
Java	34,763	9.92%	3.47%
Rust	23,558	6.72%	2.35%
Go	23,556	6.72%	2.35%
C	23,319	6.65%	2.33%
Shell	12,066	3.44%	1.20%
SQL	11,942	3.41%	1.19%
Total Code	350,419	100.00%	34.99%

2. Natural Languages Breakdown (Reasoning, Prose & Tools - 651,053 Tokens)

Language	Reasoning Tokens	Prose / Text Tokens	Tools Tokens	Total Language Tokens	Corpus Share
English (en)	51,930	51,907	31,123	134,960	13.48%
Spanish (es)	37,052	37,350	22,744	97,146	9.70%
Chinese (zh)	37,033	37,037	22,252	96,322	9.62%
French (fr)	29,649	30,150	17,779	77,578	7.75%
Portuguese (pt)	18,524	19,055	11,372	48,951	4.89%
German (de)	18,612	18,733	11,215	48,560	4.85%
Russian (ru)	14,903	15,129	9,270	39,302	3.92%
Japanese (ja)	14,882	14,931	9,171	38,984	3.89%
Arabic (ar)	11,148	11,224	6,809	29,181	2.91%
Italian (it)	7,467	7,770	4,860	20,097	2.01%
Korean (ko)	7,441	7,719	4,812	19,972	1.99%
Subtotal	248,641	251,005	151,407	651,053	65.01%

All document boundaries and token allocations were verified against the official native tokenizer to guarantee strict zero-deviation quotas across all language partitions.

MoE Expert Coverage & Physical Verification

In an MoE architecture with 256 experts per layer and top-k=8, token observations are distributed across 10,240 independent routing experts (30,720 distinct projection matrices):

Metric	Measured Value	Standard Required
Total Expert Projection Matrices	30,720 (40 layers x 256 experts x 3 FFN projections)	30,720
Observed Expert Matrices	30,720 (100.0%)	100.0%
Unobserved Experts	0	0
Minimum Observations per Expert	440 rows	Greater than 0
Maximum Observations per Expert	361,813 rows	Finite F32
Average Observations per Expert	31,296.0 rows	Finite F32
Validation Status	`VALIDATED_OBSERVATIONS`	Validated

Artifact Verification

File Name	File Size	SHA256 Checksum	Format
`Sage-A1.5-VAL-1M-I-Matrix.gguf`	384,359,808 bytes	`634336253dedbebb63643dc6cd1b51821c2132187c45c5a044a5d43d6c8b813d`	GGUF v3

Support & Independent Research

All compute hours and hardware allocations are funded 100% out-of-pocket to bring high-quality open-source quantizations and calibration assets to the community.

If these builds help you in your day-to-day workflow, research, or development, consider supporting on Ko-fi. Every contribution directly sustains future cloud compute hours and supports my university studies.