RWKV-7 G1L 7.2B — rwkv7-g1l-7.2b.knot

A Q8_0 + F32-island knot container of RWKV-7 "G1L" (GooseOne) 7.2B base weights, packed by Forkjoin.ai's sovereign encoder from the pinned RWKV org safetensors conversion. This is a base model suitable for post-training and fine-tuning, not an instruction-tuned assistant.

This is not a transformers, safetensors, or GGUF checkpoint. The .knot container is a name-addressed tensor store read by the sovereign knot runtime. Loading it with from_pretrained will not work.

Model identity

field	value
Model	rwkv7-g1l-7.2b
Family	RWKV-7 G1L (GooseOne), released 2026-10-31 by BlinkDL
Base model	RWKV/RWKV7-G1l-7.2B-20261031
Pinned source revision	100541a755612fa970f77e04c17d3b25ff7d9d29
Nominal parameter count	7.2B
Training context	34,304 tokens
Vocabulary	65,536 (RWKV World)
Artifact	rwkv7-g1l-7.2b.knot
License	Apache-2.0 (inherited from the base model)

Architecture

RWKV-7 is a recurrent linear-attention architecture: token-by-token constant state, no growing KV cache, so decode cost is independent of sequence position within the 34,304-token training context.

• L32-D4096, 64 heads x 64, gate 480
• Width / depth: 4,096 hidden, 32 blocks, 65,536-token World vocabulary, 64 heads x 64, gate inner dim 480, FFN 16,384, low-rank W/A/V 128 / 128 / 96
• This rung carries 1,062 tensors (386 Q8_0, 676 F32); the inventory is fixed by config.

Artifact and container format

field	value
Bytes	7,755,945,289
SHA-256	f5bc74acbc612b11b77764a2b87ce8d557ebc490024ffbe1965d6ee9deea137f
Tensors	1,062
Payload bytes	7,755,857,920
Header bytes	87,369

Container layout (verified by parsing this artifact's header):

offset 0    magic   "KNOT" (4 bytes)
offset 4    u16     container version = 1
offset 6    u32     header length in bytes
offset 10   ...     UTF-8 JSON header: name, arch, config, tensors{offset,length,type}
offset 10+H ...     3-byte alignment trailer ending in codec marker 0x20
offset 10+H+3 ...   payload

The JSON tensor table names each tensor and pins its byte offset, length, and storage type. Tensors of 65,536 elements or more are stored Q8_0 (34 bytes per 32 values: one f16 scale plus 32 int8); smaller tensors, and the forced-F32 rwkv_w1/rwkv_w2 islands, are stored F32. Q8_0 is the default codec; the trailer's 0x20 marker identifies it. Because the encode policy is fixed, the byte count is a pure function of shapes — token count, batch size, and runtime state do not change the artifact.

Provenance

1. Upstream base weights: BlinkDL/rwkv7-g1 (Apache-2.0), published 2026-10-31.
2. Official RWKV org safetensors conversion: RWKV/RWKV7-G1l-7.2B-20261031 at revision 100541a755612fa970f77e04c17d3b25ff7d9d29. The converted weights differ from the .pth only by metadata (<= 0.07 MB) and are the same tensors.
3. Forkjoin.ai sovereign encode: encode-ssm-models.py with the additive RWKV-7 G1L naming map and the low-rank orientation fix, then a structural gate (verify-rwkv7-knot.py) that refuses missing tensors rather than silently zero-filling them.
4. Landed on R2 at https://r2.edgework.ai/models/rwkv7-g1l-7.2b.knot; this HF artifact is the byte-identical object (same size and SHA-256).

License and attribution. The base model is Apache-2.0 and this derivative keeps those terms; the base model's attribution and upstream links are reproduced here. No additional restrictions are added. If you redistribute this artifact, keep this attribution and the Apache-2.0 license.

Admission: two-sided probe receipt

A landed artifact is not an admitted model. The admission gate is a two-sided probe on the landed knot, using RWKV World token ids:

side	expectation	observed
accept	[6699, 51128, 4706, 44312, 4600] -> 37138 (" Paris")	37138 PASS
reject	22590 must not win the argmax	22590 did not win

Dense single forward over the 5-token prompt: 64.89 s (dedicated 3x median). This is an admission witness, not a throughput benchmark.

Measured serving performance

Single-stream decode, greedy, kitchen-LAN fat station (durata, 8 threads), operator-recorded. Full ladder for context:

rung	LAN tok/s (8 threads)	repo
1.5B	3.604	forkjoin-ai/rwkv7-g1l-1.5b
2.9B	2.002	forkjoin-ai/rwkv7-g1l-2.9b
7.2B	0.890	forkjoin-ai/rwkv7-g1l-7.2b
13.3B	0.687	forkjoin-ai/rwkv7-g1l-13.3b

The 13.3B lane was served from a separate station (serata, 10.0.0.189:9104); 7.2B/13.3B are full-resident / R2-Range lanes, while 1.5B/2.9B fit the layer-sharded fleet envelope.

Do not read these as portable benchmarks. They are measured on owned LAN hardware, single stream, and vary with contention and the tensor cache. The repository receipts carry the method, repetition counts, and raw samples; a slope-based repeat measurement on the same 1.5B station is also recorded and differs under contention.

Using this artifact

The knot runtime is part of Forkjoin.ai's sovereign inference stack; see open-source/gnosis/distributed-inference/RWKV7_G1L_ONBOARDING.md for the container and admission contract, and the receipts under open-source/aether/docs/model-cards/ for the raw measurements.

Ladder

• rwkv7-g1l-1.5b
• rwkv7-g1l-2.9b
• rwkv7-g1l-7.2b
• rwkv7-g1l-13.3b

Provenance note

These artifacts were encoded, admitted, and published with sovereign local tooling. No platform-paid AI service was used at any stage.