Ricursive releases Ripple-65B open weights: $32M model hits US open-weight SOTA on TerminalBench 2.1
DZhang50 · x · 2026-11-05
Ricursive Intelligence released Ripple-65B Preview, a from-scratch 65B dense LLM (80 layers, d=5120, built on Olmo 3 roots) claiming US open-weight SOTA on agentic coding: 80.5 on TerminalBench 2.1, beating much larger models like Reflection Beam-501B (80.1), Muse Spark Open-816B (76.4), Inkling-975B (63.8) and Nemotron 3 Ultra-550B (56.4).
Key facts:
- Pretrained on 12T tokens, up to 1M context, full R&D cost $32M (roughly 1/10 of frontier model budgets)
- Architecture: alternating 4 SWA-1024 (RoPE) : 1 global NoPE layers, 80q/8kv GQA, σ-gated attention, QK-norm, value residuals from embeddings
- Tiny KV cache (64 KiB/token, 16 global layers only); NVFP4 quantization at 7.1 bits/weight shrinks it to 57 GB — runs on a DGX Spark
- Serves as the general-purpose variant of the proprietary model powering Ricursive's AI chip-design platform
Weights are out under Apache 2.0 on Hugging Face, Together AI, GMI Cloud, and OpenRouter, with Fireworks coming later.
Related event: Ricursive Open-Sources Ripple-65B, Claims US Agentic Coding SOTA(16 posts)→
More from Models
- Ripple claims frontier-level open-weight model trained in 32 minutes at one-tenth cost — annadgoldie · 2026-11-05
- Variable names matter: safety-flavored naming can push LLMs into over-cautious mode — jessi_cata · 2026-11-05
- Developers report Grok API instability: 13.58% error rate and 13.6s p90 latency in dashboards — ChrisUniverse · 2026-11-05
- Ramp's post-trained Qwen beats Opus on finance QA; 10,000-agent swarm finds 7 critical bugs — AI Engineer · 2026-11-05
- Vals AI: four of seven frontier models keep 80%+ scores at lowest reasoning effort for half the cost — ZacharyHuang12 · 2026-11-05
- Sailing sim race shows Chinese models lag frontier: "cheap tokens ≠ cheap agents" — shekitup · 2026-11-05