RWKV-7 G1j released: pure RNN architecture gets much better at agents and coding
jeremyphoward · x · 2026-09-02
BlinkDL has released RWKV-7 G1j weights, claiming significant gains on agent, coding, and STEM tasks. RWKV is a 100% attention-free RNN that combines Transformer-level performance with linear time, constant space (no kv-cache), fast training, and unlimited context length, and is a Linux Foundation AI project. The ecosystem includes 7B/13B demos, GGUF/Ollama weights, RWKV-PEFT (fine-tune 7B with 9GB VRAM), the Albatross inference stack (7B fp16 at 10250+ tps on a 5090), and 800+ community projects.
More from Models
- Grok Image Generation Fail: Predicts Student Will Become Street Mascot — burkov · 2026-09-02
- Hands-on with Claude 5.1: The Strongest Coding Model That Speaks Human — danshipper · 2026-09-02
- claude-fable-5 resellers offer 64% off: $3.60 in, $17.99 out per Mtok — const_reborn · 2026-09-02
- Tokenomics 101: understanding input, output, and cached token pricing in the AI era — Aizkmusic · 2026-09-02
- Fable 5.1 First Impressions: High Pricing, Less Optimization, Shift to Complex Tasks — Aizkmusic · 2026-09-02
- Anthropic accused of ignoring Gemini, Grok, and Chinese open source — himanshustwts · 2026-09-02