HaloWang/rwkv-weights

RWKV Weights

This repository distributes final consumer artifacts with immutable paths and SHA-256 identities. Source checkpoints for the initial G1i cohort come from BlinkDL/rwkv7-g1 at revision 159cf82ed0f18eb9dcd92c388688bec11024895e.

Layout

artifacts/<family>/<task>/<size>/<backend>/<basename>
manifests/<family>/<task>/<cohort>.json

Generation, quantization, context length, converter or SDK revision, platform, accelerator, SoC, ABI, architecture, and batch mode are identity metadata, not directory levels. They remain in the immutable manifest and, when needed to distinguish bytes, in the globally unique basename.

Use index.json for discovery and the referenced immutable manifest for sizes, digests, source lineage, recipes, and consumer-acceptance status. Binary paths and cohort manifests are immutable after publication.

G1i cohorts

The completed Apple and WebRWKV publication contains eleven exact binaries:

• WebRWKV NF4 1.5B, 2.9B, 7.2B, and 13.3B artifacts accepted through the exact macOS and Windows consumer paths recorded by the publisher
• MLX INT6 1.5B, 2.9B, 7.2B, and 13.3B artifacts accepted on macOS
• compiled Core ML INT4 1.5B, 2.9B, and 7.2B artifacts accepted on macOS

The separate 2026-08-10 QNN cohort adds 21 exact .rmpack files:

• eleven G1i 1.5B A16W8 packs for Android Snapdragon and Windows Snapdragon
• ten G1i 2.9B A16W4 LPBQ block-32 packs for Android Snapdragon and Windows Snapdragon

All 21 QNN files passed byte-size, SHA-256, package-header, member-table, and byte-range verification. The generated 1.5B and 2.9B Snapdragon 8 Gen 3 packs also passed the retained Xiaomi 14 HTP runtime smoke. Other target-specific packs remain distribution artifacts with structural verification; publication does not imply that every listed SoC received exact-device runtime acceptance.

A later dedicated QNN cohort adds the exact G1i 7.2B Snapdragon 8 Gen 3 A16W4 LPBQ block-32 five-chunk pack. It passed package structure, local/device SHA-256 identity, and native Xiaomi 14 QNN HTP V75 execution with one warm-up and three measured 384-prompt-token plus 64-decode-step repetitions. This is native runtime acceptance, not a claim about the application's user interface.

The separate 2026-08-10 llama.cpp cohort adds four exact .gguf files from BlinkDL/rwkv7-g1 revision ede85bf8ab2e59aff7d7ca909fbbc73317866d89:

• G1i 1.5B Q6_K
• G1i 2.9B, 7.2B, and 13.3B Q4_K_M with Q6_K embedding/output tensors

All four GGUF files passed full SHA-256 revalidation and retained fixed pp512/tg128 llama.cpp acceptance on Windows CPU and Xiaomi 14 Android CPU. These results are device- and protocol-scoped. The 13.3B Android run proves functional loading and generation only; it was extremely slow and is not a performance recommendation.

G1i 0.1B and 0.4B have no source checkpoint in the frozen source revision and remain explicit source-unavailable rows. The generated Core ML 13.3B package is not included because final consumer acceptance has not completed. The original QNN cohort did not contain 7.2B or 13.3B; the later accepted 7.2B pack is now published separately, while no exact final 13.3B QNN pack was found. No HarmonyOS artifact is included. Publication here does not change any application's live model catalog.

Verification

Treat a filename or mutable URL as discovery only. Require publicationState: published, then compare the downloaded byte size and SHA-256 with the applicable immutable manifest:

• manifests/rwkv7/chat/g1i-20260810.json for Apple and WebRWKV artifacts
• manifests/rwkv7/chat/g1i-qnn-20260810.json for QNN artifacts
• manifests/rwkv7/chat/g1i-llamacpp-20260810.json for GGUF llama.cpp artifacts
• manifests/rwkv7/chat/g1i-qnn-7p2b-20260810.json for the dedicated G1i 7.2B Snapdragon 8 Gen 3 QNN artifact