Ricursive drops Ripple-65B: Apache-2.0 coding weights, runs locally

A 65B dense model the lab says tops US open-weight agentic coding — scores still unaudited. Plus Kimi K3.1 weights, an AI-image detector, free OpenAI credits.

Nowline NOV 5 3:00 AM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • Open weights, Apache 2.0, downloadable now

    Ricursive Intelligence released Ripple-65B Preview — a 65B dense, 80-layer coding model — ungated on Hugging Face under Apache 2.0, so you can ship it commercially with no agreement. The BF16 checkpoint is ~130 GB (needs a 192 GB-class accelerator), an experimental NVFP4 build is 57 GB and fits a 128 GB unified-memory box, and there's a 4B DFlash2 draft model for speculative decoding.

  • 'US open-weight SOTA' — with an asterisk

    Ricursive claims 80.52 on Terminal-Bench 2.1, calling it the top US open-weight model for agentic coding and ahead of rivals up to ~975B params. But the scores are vendor-reported, run under Ricursive's own harness, and unreproduced — Artificial Analysis has no entry yet. It's tuned for long-horizon terminal work and thinner elsewhere (MCP Atlas 71.7 vs ~89; SciCode last of five), so treat it as a promising shell-agent model, not a settled champion.

  • 1M context on the card, 16K on the only hosted API

    The model card advertises up to 1M tokens, but the first hosted endpoint serves just 16,384 (≈14.7K output) at $0.20 / $0.60 per million in/out. If you actually need the long window, self-hosting is the only path today — and that's where the KV-cache memory bill lands.

  • Update: Kimi K3.1 Pro open weights are live

    After yesterday's 'due today,' Moonshot's Kimi K3.1 Pro is now on Hugging Face: a 2.8T-param MoE (~104B active) with a 1M-token context and a MoonViT-V2 vision encoder, shipped as MXFP4 weights. Another frontier-scale open model to self-host — if you've got the silicon for it.

  • Pangram ships an AI-image detector with browser add-ons

    Pangram Image 1.0 (plus a faster Image Flash 1.0) claims 99.96% accuracy across 1,200 images from 12 frontier generators and 98.7% on the DF26 video benchmark. It drops the minimum image size from 512 to 256 px and adds Chrome and Firefox extensions — useful if you're wiring provenance or moderation into a product.

  • OpenAI is giving Pro/Business subscribers up to $500/mo in API credits

    A quieter changelog item worth claiming: eligible ChatGPT Pro and Business subscribers can now pull up to $500 per month in API credits, activated by linking an API org you own. If you already pay for ChatGPT, that's a meaningful chunk of free build budget each month.