borkiss/Nanbeige4.2-3B-WebGPU

Nanbeige4.2-3B · browser WebGPU artifacts

Run locally in your browser · Source and reproducible tooling · Report

Custom WGSL artifacts for the original Nanbeige4.2-3B, source revision 3384e426066d1a49c3aea90a7190b81260a6533f. This is an independent browser port, not an official Nanbeige release. The repository is not directly loadable by AutoModel; use the linked app or its custom runtime.

Artifact directory	Download bytes	Use
q4g32-e4-gram8192-emse-packed	2,345,711,616	Default compact profile; calibrated Q4 linears and Q4 MSE embedding
q4g32-e8-gram8192-packed	2,600,908,800	Calibrated Q4 linears, Q8 embedding control
q4g32-e8-packed / q4g32-e8	2,600,908,800	Original round-to-nearest control, packed / individual files

The compact model was first published at revision 8d36b29983c2ad1465be29fa51ebace46b91dc62; the calibrated Q8 embedding profile at 32adccbdde7adcab3f253678c414245a8b1072d9. The app pins immutable revisions. Each manifest gives file SHA256 values, tensor shapes, byte offsets, scales and the complete source config. Packed bundles are at most128MiB; GPU tensor bindings fit the128MiB WebGPU baseline.

The runtime shares22 physical blocks across exactly two passes, with44 independent KV slots and a final RMS normalization after each pass. It preserves48 query heads,8 KV heads and128-dimensional full RoPE. No layer pruning or reduced recurrence is used. Browser context is intentionally capped at2048 (default512) to bound memory; this is a deployment limit, not the source model's context limit.

Quantization uses signed Q4 groups of32 with FP16 scales and byte-sliced nibbles. A group covariance reconstruction objective calibrates linears on8192 WikiText-2 train tokens, observing both recurrent visits. It is a bounded group reconstruction experiment, not full GPTQ/AWQ. Compact embedding uses per-group FP16 scale search minimizing weight MSE. Quantization and all full-weight processing ran on an RTX A6000; none ran on the development Mac.

Small screening evaluation: same4096 WikiText-2 test token IDs,4080 scored tokens,16 independent256-token windows:

Model	Perplexity	Six-prompt mean KL vs source	First-token agreement
Source FP32	41.774	0	6/6
Original Q4g32/Q8 embedding	48.828	.07496	6/6
Calibrated Q4/Q8 embedding	45.237	.06777	6/6
Compact calibrated Q4/Q4 embedding	44.859	.07180	6/6

These are limited screening measurements used to choose among candidates, not a full unbiased benchmark or a claim that quantization improves the source. The separate tiny multilingual corpus slightly favors Q8 embedding. Raw losses, prompts, logits, calibration provenance, other candidate failures and scripts are in the Space's results/ and source/ directories.

WebGPU is required. Desktop Chrome and Safari are measured on M4; physical iPhone15ProMax, Pixel10a and iPadM2 acceptance remains unverified. Device memory and browser limits can still prevent a3B-class model from loading. No inference server receives prompts. Hugging Face serves static files and weights; downloads require network access.

Source model weights are Apache-2.0. Runtime source has separate MIT notices for reused kernels. Please retain the original model attribution and consult its technical report.