One-line: ⚠ no published benchmarks — agent-loop suitability unknown. Official BlinkDL RWKV-7 “Goose” G1k weight family (1.5B/2.9B/7.2B/13.3B), dropped 2026-09-30; Apache-2.0, attention-free constant-state RNN, training context raised to 25.6K (G1j was 16K).
Coding benchmarks
Benchmark
Score
Source
HumanEval
—
none published
HumanEval+
—
none published
MBPP
—
none published
LiveCodeBench
—
none published
Aider polyglot
—
none published
SWE-bench Verified
—
none published
SWE-bench Multilingual
—
none published
Agentic / tool-use benchmarks
Benchmark
Score
Source
BFCL
—
none published
τ-Bench
—
none published
ToolACE
—
none published
GAIA
—
none published
Reasoning (context only)
Benchmark
Score
Source
MMLU / MMLU-Pro
—
none published
GPQA Diamond
—
none published
MATH
—
none published
AIME 2025
—
none published
Efficiency
Active params: total = all (dense/recurrent; no MoE). Sizes 1.5B (L24-D2048) / 2.9B (L32-D2560) / 7.2B (L32-D4096) / 13.3B (L61-D4096)
Context length: 25,600 base (up from G1j’s 16,384; state-extrapolatable beyond — .pth filename *ctx25600)
Inference: constant recurrent state, constant work per token (no KV-cache growth) — BlinkDL cites ~145 tok/s for 7.2B fp16 bsz1 on RTX 5090; bundled TileLang/PyTorch inference/ bundle + Transformers 5.15+ integration (trust_remote_code)
What makes it notable
Third monthly-ish Goose family drop (G1i 08-05 → G1j 08-31 → G1k 09-30), and the first with an official Transformers release package: bundled configuration_rwkv7.py/modeling_rwkv7.py, exact linear-time tokenizer, SFT/LoRA workflow (TRL 1.10+ SFTTrainer), and an isolated inference bundle with TileLang acceleration. Ships a chat template with thinking prefix and strict tool-call prompt support (“Return only a JSON function call” / <tool_call> formats), so it is chat/tool-template-capable despite being a base checkpoint (“pretrained with web, code, synthetic, instruction, chat, and reasoning data”). Zero published benchmarks — no SWE/HumanEval/LiveCodeBench/Aider/BFCL numbers, so agent-loop quality is unverified. Custom architecture (model_type: rwkv7, Rwkv7ForCausalLM) — stock transformers 5.15+ handles it but the bundled remote code is the recommended path; GGUF support lives in the community collection (shoumenchougou/rwkv7-gxx-gguf), typically lagging the raw drop. Surfaced via per-org author= sweep + one-repo-many-version sibling check (raw weights at BlinkDL/rwkv7-g1, dated 20260930): dropped today at 0 downloads, invisible to the createdAt/downloads/lastModified streams.