Research Vault

BlinkDL/RWKV-7 G1 (GooseOne) — G1i + G1j

Properties1
tagshuggingface, text-generation, coding, agentic, rnn, attention-free, recurrent

3 min read

BlinkDL/RWKV-7 G1 (GooseOne) — G1i + G1j

One-line: ⚠ no published coding/agent benchmarks — agent-loop suitability unknown. Attention-free pure-RNN RWKV-7 “Goose” architecture; G1i family (1.5B/2.9B/7.2B/13.3B, ctx 16384) dropped 2026-08-05, then a new G1j family (same sizes, ctx 16384) dropped 2026-08-31; headline feature is constant VRAM (no KV cache) — potentially the cheapest long-context local runner, unproven for coding agents.

⚠ No coding or agentic benchmarks published (lab or third-party) as of 2026-08-08. The BlinkDL card and the official safetensor card (RWKV/RWKV7-*-20260805) both ship prompt/tool-call templates but no eval tables. Agent-loop suitability is untested. Community (Zhihu) commentary: “code and STEM still behind Qwen3.5, expected to catch up by year-end” — qualitative, not a number.

Release event: the BlinkDL/rwkv7-g1 repo is old (created 2025-03-07) but the G1i weights were uploaded 2026-08-05 (rwkv7-g1i-{1.5b,2.9b,7.2b,13.3b}-20260805-ctx16384.pth); community GGUF quants (shoumenchougou/RWKV7-G1i-*-GGUF) and official safetensor releases (RWKV/RWKV7-*-20260805, transformers 5.15+ Rwkv7ForCausalLM) appeared 08-05/08-06. On 2026-08-31 a further G1j family landed in the same repo (rwkv7-g1j-{1.5b,2.9b,7.2b,13.3b}-20260831-ctx16384.pth) — this run’s trending signal. G1i/G1j are the newest data versions — the card says “always use latest models, they are better at everything.”

Coding benchmarks

BenchmarkScoreSource
HumanEval——
HumanEval+——
MBPP——
LiveCodeBench——
Aider polyglot——
SWE-bench Verified——
SWE-bench Multilingual——

Agentic / tool-use benchmarks

BenchmarkScoreSource
BFCL——
τ-Bench——
ToolACE——
GAIA——

Reasoning (context only)

BenchmarkScoreSource
MMLU / MMLU-Pro——
GPQA Diamond——
GSM8K77.3 (older G1a4-2.9B)RWKV wiki, community-reported
AIME 2025——

Language-modeling PPL (Uncheatable Eval, community, older G0a3-7.2B): avg 7.222 vs Qwen2.5-7B 7.323 (lower=better) — but this is a different benchmark class (PPL on raw corpora incl. github cpp/python), not a coding task eval.

Efficiency

  • Active params: 1.5B / 2.9B / 7.2B / 13.3B (dense — every param active; pure RNN, no MoE)
  • Context length: 16,384 tokens (G1i; older G1d was 8192)
  • VRAM (fp16 rough): 1.5B ≈ 3 GB · 2.9B ≈ 5.8 GB · 7.2B ≈ 14.4 GB · 13.3B ≈ 26.6 GB
  • Constant VRAM/speed — no KV cache, memory does not grow with context length (unlike attention models)
  • Measured (RWKV wiki, 2.9B G1): fp16 5.52 GB (rwkv pip) / 89 tok/s llama.cpp CUDA; Q8_0 3.47 GB / 110 tok/s; int8 3.9 GB; nf4 2.4 GB
  • Card claim: 145+ tok/s RWKV-7 7.2B fp16 bsz1 on RTX 5090 (Albatross engine), constant VRAM

What makes it notable

RWKV-7 is the flagship attention-free recurrent architecture (paper arXiv:2503.14456) — linear-time, constant-state, 100% RNN. G1i is a fresh pretrained family with 16k context, and for the first time ships as official Transformers-format safetensors (Rwkv7ForCausalLM, transformers 5.15+) with a chat template that supports system/multi-turn/thinking and strict model-generated tool-call prompts (function-call docs: temp 0, topp 0, penalty 0, G1f and newer). For a local runner this is the cheapest long-context option: no KV cache means memory stays flat on long agentic sessions, and 13.3B runs in ~27 GB fp16 (or ~7 GB nf4-class quant). The catch: zero published coding/agent benchmarks — the “no KV cache + tool-call template” pitch is unverified for actual agent loops, and community commentary puts its code/STEM below Qwen3.5 for now. GGUF builds exist for llama.cpp; the raw .pth needs the RWKV pip package / RWKV-LM.

See also

Graph View

Backlinks