Groot-65B-A6B-261023 VAL 1.25M I-Matrix GGUF

The complete 611-window artifact passed the model-specific statistics checks. These checks establish integrity and observation coverage; they do not measure quantization quality. Calibration infrastructure and future releases can be supported through Ko-fi.

Technical Overview

Importance statistics for GrootTeam/Groot-65B-A6B-261023, collected for subsequent GGUF quantization. The selected coverage is causal text inference. No perplexity, downstream benchmark or measured quantization-quality claim is included.

The collector is official llama.cpp b11512, commit a11f57ba93797579a5d1855ee216a31f10242676, with Flash Attention explicitly enabled. The source required a fixed-width correction of its compression-ratio metadata: the 12 full-attention trunk layers use ratio 4, matching the pinned official configuration and converter. Weight payloads are unchanged.

Parameter	Specification
Model Repository	`GrootTeam/Groot-65B-A6B-261023`
Model Revision	`d8fd3c7ce58b29a8413e27f75f14114a32387153`
Model Architecture	`qwen4exp`: 48 causal layers, comprising 36 linear-attention and 12 full-attention layers; 256 routed experts per layer; top-k=10; expert intermediate size 640
Additional Components	One MTP block exists in the source; MTP and vision are outside this text-only calibration
Source Repository	`mradermacher/Groot-65B-A6B-261023-GGUF`
Source Revision	`440ec2ddf1724c1db5509e1d83e9efa27df37b41`
Corrected Source SHA256	`fb6a3802dfb971f498a3473c8b9d7420136a4961edf38c6adb5d96a3accc9358`
Target Tokens	1,250,000 target tokens; 611 x 2048 = 1,251,328 planned positions
Prepared Corpus	Universal v5; `universal-pool-v3`; native accounting of the exact processed prefix
Execution Backend	CUDA; NVIDIA RTX PRO 6000; Flash Attention on
Execution Outcome	All 611 windows completed and the matrix passed validation. A duplicate persistence-probe copy caused the runner to exit with an error afterward; the bucket artifact was recovered and independently revalidated.
Execution Flags	`--process-output --no-ppl --no-escape --parse-special -c 2048 -b 2048 -ub 2048 --chunks 611 -ngl 99 -fa on --device CUDA0 -t 23 --load-mode mmap --lazy-mode on --fit off`

Macro Component Distribution (1,251,328 Positions)

The prepared prefix follows a deterministic 30/20/35/15 target mix across 12 natural languages and 10 programming languages. English and Spanish each receive weight 20/120 within each natural-language category; the other ten languages each receive 8/120. The available pool and the assembled tail are larger than the 611-window processing budget.

Component	Target Share	Measured Share	Effective Tokens	Sub-categories / Scope
Code	30.0%	30.03%	375,790	10 programming languages: Python, JavaScript, TypeScript, Java, C++, C, Go, Rust, SQL and Shell
Reasoning	20.0%	19.92%	249,315	12 natural languages; mathematical and structured reasoning
Prose / Text	35.0%	34.89%	436,597	12 natural languages; prose, encyclopedic and technical material
Tools / Agentic	15.0%	15.15%	189,626	12 natural languages; tool schemas, calls, results and multi-turn transcripts
Total Positions	100.0%	100.0%	1,251,328	611 chunks of 2048 tokens

Detailed Token Breakdown by Language & Category

1. Code Breakdown by Programming Language (375,790 Tokens)

Programming Language	Effective Tokens	Category Share	Corpus Share
Python	100,203	26.66%	8.01%
JavaScript	62,332	16.59%	4.98%
TypeScript	37,731	10.04%	3.02%
Java	37,602	10.01%	3.00%
C++	37,440	9.96%	2.99%
C	25,176	6.70%	2.01%
Go	24,936	6.64%	1.99%
Rust	24,905	6.63%	1.99%
SQL	12,739	3.39%	1.02%
Shell	12,726	3.39%	1.02%
Total Code	375,790	100.00%	30.03%

2. Natural Languages Breakdown (Reasoning, Prose & Tools - 875,538 Tokens)

Language	Reasoning Tokens	Prose / Text Tokens	Tools Tokens	Total Language Tokens	Corpus Share
English (en)	41,422	72,967	31,444	145,833	11.65%
Spanish (es)	41,421	72,467	31,337	145,225	11.61%
Chinese (zh)	16,628	29,122	12,788	58,538	4.68%
Hindi (hi)	16,633	29,086	12,464	58,183	4.65%
Arabic (ar)	16,577	29,124	12,864	58,565	4.68%
French (fr)	16,673	29,139	12,601	58,413	4.67%
German (de)	16,572	29,019	12,742	58,333	4.66%
Portuguese (pt)	16,704	28,995	12,670	58,369	4.66%
Japanese (ja)	16,722	29,103	12,441	58,266	4.66%
Russian (ru)	16,713	29,207	12,884	58,804	4.70%
Korean (ko)	16,661	28,993	12,601	58,255	4.66%
Italian (it)	16,589	29,375	12,790	58,754	4.70%
Subtotal	249,315	436,597	189,626	875,538	69.97%

Counts describe the exact prepared native-token prefix; percentages are rounded. Full HF/native token IDs differ for some Unicode normalization sequences, including Hindi nukta forms. The unchanged selector verifies native reconstruction of the source bytes and derives quotas from native IDs. The full native ID sequence is checked again in Linux. This is not a claim of identical HF/native tokenization or zero-deviation quotas.

MoE Expert Coverage & Physical Verification

The causal trunk contains 12,288 routing expert slots (48 x 256), represented by 36,864 per-expert projection matrices (48 x 256 x 3 FFN projections). These are stored as 144 grouped expert tensors. Each routed tensor must total positions x top-k: 12,513,280 observations for a complete 611-window run. Individual experts see only their routed tokens.

Metric	Measured Value	Standard Required
Total Expert Projection Matrices	36,864 declared by the causal inventory	48 x 256 x 3
Observed Expert Matrices	36,864 (100.00%)	Report actual observations
Unobserved Expert Matrices	0	Report zeros as partial coverage
Minimum Observations per Expert	519 rows	Nonnegative integer counts; zeros explicitly identified
Maximum Observations per Expert	592,527 rows	Finite F32 statistics
Average Observations per Expert	48,880.0 rows	Counts verified against the routed total
Required Causal Tensor Entries	831 in the reviewed contract	Every required entry present with the correct dimensions
Validation Status	`VALIDATED_AGAINST_AGENT_CONTRACT`	Exact 611 x 2048 volume, complete F32 statistic pairs and finite raw values

The contract also records 48 diagnostic routers and 377 exclusions. Exclusions include MTP, embedding gathers, norms, recurrent/depthwise operations, and the global output_hc_* projections that the b11512 collector name filter does not collect. No importance coverage is claimed for these exclusions. Low expert counts are diagnostic and are not a quality guarantee.

Artifact Verification

File Name	File Size	SHA256 Checksum	Format
`Groot-65B-A6B-261023-VAL-1.25M-I-Matrix.gguf`	593,456,000 bytes	`bab269361aedf3f97e2fbe95bbc4efa389fed222b33b43f5c5b31262797befe1`	GGUF v3

Support & Independent Research

Contributions help fund calibration runs, maintenance of reproducible tooling and future model releases. If these artifacts are useful to your work, you can support their continued development on Ko-fi.