MiniCPM5-2B WebGPU Q4 G32

Activation-calibrated Q4 group-32 weights for the purpose-built MiniCPM5-2B WebGPU runtime. This is a modified quantized distribution of the official OpenBMB MiniCPM5-2B final checkpoint, not a separately trained model.

Format and use

artifact/manifest.json describes 43 content-addressed weight chunks and one tokenizer container. Download files relative to that manifest and verify their declared SHA256 hashes before use. Pin the repository commit rather than main for reproducible application deployments.

  • Private minicpm5-webgpu-artifact version 2, row_plane_v2.
  • Signed 4-bit weights with one FP16 scale per 32 weights; norms are FP32.
  • Weight payload: 1,416,273,920 bytes; tokenizer: 9,988,682 bytes.
  • This is not GGUF or a Transformers/safetensors checkpoint. It requires the matching custom WebGPU runtime.
  • INT8 projection activations (W4A8) are a runtime setting, not stored weights. KV remains FP16; these files do not prescribe KV quantization.

Provenance

Original checkpoint: openbmb/MiniCPM5-2B@42433e90dad95123ab87729fd7011e7110629556

Original BF16 safetensors SHA256: 14fb8e7f0a18d53d1f239773758bf581cee7e456a4523a54622c3a245b64402c

Conversion reads that original BF16 source directly, without requantizing another low-bit checkpoint. Activation-second-moment-weighted scale search uses 12 independent authored calibration fixtures and 1,523 samples per projection input tensor. It is diagonal importance weighting, not full-Hessian GPTQ or AWQ.

Calibration SHA256: cb802025f2793e5bf04aa458af238d869cc2b398110e2eb93c90b9544239ac4d

Artifact identity: d3eaa8777abd8eaaf490e687de67b766e7445d9b7e8977fa2672e7f0c6a1533f

provenance/weight-verification.json records independent re-encoding and exact comparison of every byte in all 255 GPU weight buffers. Calibration moments and quantization statistics are included alongside it. Raw BF16 weights and experimental precision variants are not included in this repository.

Limitations and licence

Quantization changes model outputs and does not establish BF16 quality parity. Small held-out task and corpus-prefix checks are regression evidence, not a comprehensive evaluation or a guarantee of identical behaviour.

Upstream model and tokenizer: OpenBMB/ModelBest, Apache-2.0. The original authors' licence applies to this modified distribution; see LICENSE. This repository is not an official OpenBMB release.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for leventkoldas/MiniCPM5-2B-WebGPU-Q4-G32

Quantized
(139)
this model

Space using leventkoldas/MiniCPM5-2B-WebGPU-Q4-G32 1