⏰ OpenRouter Limited-Time Free Models

Several models remain available on OpenRouter’s limited-time free tier:

  • Poolside: Laguna M.1 (free — expires today, July 27) — 262K context, $0.20/M tokens input / $0.40/M tokens output
  • OpenAI: GPT-5.3 Chat (free until August 10) — 128K context, $1.75/M tokens input / $14/M tokens output
  • OpenAI: GPT-5.2 Chat (free until August 10) — 128K context, $1.75/M tokens input / $14/M tokens output

Two new Claude Opus 5 variants appeared on OpenRouter this cycle:

  • Anthropic: Claude Opus 5 — 1M context, $5/M tokens input
  • Anthropic: Claude Opus 5 (Fast) — 1M context, $10/M tokens input

New uploads this cycle:

  • Nimbus-Labs/Nimbus-2B-GGUF — Qwen3.5-based 2B multimodal vision model in GGUF format. 576 downloads on debut.
  • Miiche/visualrl-agopd-5p0 and visualrl-opd-5p0 — Visual RL research checkpoints
  • analist/poetlm-29m-sft and poetlm-29m-pretrain — Small poetry language models trained from scratch
  • zhzh98/gpt2-dutch-mono-merge — GPT-2 fine-tuned for Dutch language
  • AeonAIRisk/npz-object-array-ace-poc — Keras-based research proof-of-concept

Top AI repositories trending this period:

  1. Nvidia’s Open Secure AI Alliance reshapes the safety landscape. Nvidia, Microsoft, and SpaceX launched an open-source AI safety coalition — notably without OpenAI, Google, or Anthropic — in response to the recent OpenAI cyber attack and Hugging Face infrastructure breach. The alliance focuses on open security frameworks rather than closed safety audits.

  2. Nvidia’s $250B financing talks with OpenAI and $500B SK Group deal signal massive infrastructure buildout. WSJ reports Nvidia is negotiating to guarantee $250 billion in data center financing for OpenAI. Separately, SK Group signed a $500 billion deal for AI infrastructure. These moves underscore the staggering capital requirements of frontier AI development and raise circular-financing concerns.

  3. Moonshot AI releases Kimi K3 as open-weight model. China’s Moonshot AI opened access to its Kimi K3 model for download, sparking debate about open-weight Chinese AI models. CNBC reports Silicon Valley and DC are increasingly obsessed with AI distillation, while the Financial Times questions the risks of deploying Chinese open models in enterprise environments.

  4. Multimodal world models gain momentum. Black Forest Labs launched FLUX 3, trained jointly on image, video, and audio as a unified world model. Alibaba released Qwen Audio 3.0 TTS with voice cloning. Microsoft expanded its MAI lineup with MAI-Image-2.5-Pro and MAI-Voice-2-Flash, deploying across Bing, PowerPoint, and Azure. These releases signal a shift from separate modality models toward unified multimodal systems.

Today’s Snapshot

  • OpenRouter models: 2 new (Claude Opus 5 variants)
  • HuggingFace uploads: 10 new models
  • GitHub trending repos: 5 AI repos
  • Limited-time free models: 4 (1 expiring today)
  • Total items collected: 21
  • Article #72 published