VBL-32M-Utility Native

31,974,240-parameter VBL utility/research release with a native Rust x86_64 runtime.

Native execution

The primary runtime is not Python/PyTorch.

• Rust x86_64
• AVX2 + FMA optimized CPU dot products
• FP32 weights
• all model weights loaded into process RAM at startup
• no mmap
• no SSD/model streaming
• no quantization
• no GPU required for runtime

FP32 neural parameter payload: 121.97 MiB.

Neural architecture

Active checkpoint architecture:

input -> embedding -> prelude -> [V16_NORM_MEAN + state-conditioned elastic bank + shared recurrent core] x R2 -> coda -> tied output

• base VBL-RC M0: 20,013,920 parameters
• connected elastic capacity: 11,960,320 parameters
• total: 31,974,240
• 73 experts
• expert width: 320 -> 256 -> 320
• top-k: 3
• router: token-causal, shares expert down.weight[0], zero additional router parameters
• selected experimental alpha: 0.75

Verification status

This repository is intentionally named Utility, not Verified-32M.

The 32M candidate's official frozen C1 result was:

583 / 853

Therefore it is NOT KR100 promoted. The canonical M0 lineage remains preserved in the 32M base weights.

Utility mode

The native executable loads the complete 32M model into RAM and first routes deterministic tasks through executable verifiers. Unsupported utility requests abstain instead of fabricating a result.

Examples:

./vbl32m --model-dir . "What is 34 + -13?"
./vbl32m --model-dir . "Is 772 even or odd?"
./vbl32m --model-dir . "Complete the sequence: 32, 33, 34, 35,"

Raw experimental neural generation:

./vbl32m --model-dir . --raw "User: What is 2 + 2? Assistant:"

Trace recurrent state convergence:

./vbl32m --model-dir . --raw --trace "User: Hello Assistant:"

Windows x86_64

The repository includes native/build-windows-x86_64.ps1.

On an x86_64 Windows machine with Rust installed:

cd native
.uild-windows-x86_64.ps1

This produces native arget eleasebl32m.exe.

VBL architecture policy

See ARCHITECTURE.json.

Features that were approved as experimental directions but were not trained/certified in this exact checkpoint are explicitly marked inactive instead of being falsely presented as learned behavior. Full-RAM residency intentionally disables predictive SSD streaming/prefetch.

Files

• model.safetensors — complete 32M FP32 state
• tokenizer.json — VBL-BPE-24K
• native_config.json — native runtime configuration and parity proof
• native/ — Rust source
• bin/linux-x86_64-v3/vbl32m — compiled Linux x86_64 AVX2/FMA binary
• ARCHITECTURE.json — active/inactive VBL architecture contract
• MANIFEST.json / SHA256SUMS — provenance