⏰ OpenRouter Limited-Time Free Models
| Model | Context | Pricing | Expires |
|---|---|---|---|
| Poolside: Laguna XS.2 (Free) | 262K tokens | Free | Jul 9 (2 days left) |
| Google: Gemini 2.5 Flash Lite Preview 09-2025 | 1M tokens | $0.10/M input → $0.40/M output | Jul 9 (2 days left) |
| Arcee AI: Trinity Mini | 131K tokens | $0.045/M input → $0.15/M output | Jul 10 (3 days left) |
Poolside’s Laguna XS.2 free tier and Google’s Gemini 2.5 Flash Lite preview are both expiring in 2 days. Arcee AI’s Trinity Mini — a compact, efficient model — also has just 3 days left on its limited-time free window. Developers looking to test these models should act soon.
New on OpenRouter: Tencent Hy3
Tencent’s Hy3 model landed on OpenRouter with 262K context at $0.20/M input, alongside a free tier for evaluation. The model packs just 21 billion parameters but reportedly matches larger rivals on key benchmarks — a strong showing for Tencent’s push into open-weight foundation models.
🤗 Trending on Hugging Face
- iamseungpil/metacot-h200-triobj-dcpo-v3 — A MetaCOT variant fine-tuned with DCPO (Direct Contrastive Preference Optimization) on H200 hardware for multi-objective reasoning. Demonstrates continued refinement of the MetaCOT line with contrastive training techniques. (⭐ 4)
- yusuf229/mam-turkish-790m — A 790M-parameter Mamba-based Turkish language model. Notable as a rare Mamba-architecture release for a non-English language, trained with TensorBoard and safetensors. (⭐ 1)
- belztjti/smg — A GLM-based model with 1,621 downloads, suggesting some community interest despite no GitHub stars yet.
- saliacoel/chars — A character-level model with 9 likes, the most popular entry in this HF cycle.
- Julian2002/PDP-Qwen3-8B-SFT-LoRA — LoRA fine-tune of Qwen3-8B using SFT, likely targeting domain-specific instruction following.
⭐ GitHub Trending: AI Edition
| Repo | Description | Stars |
|---|---|---|
| elder-plinius/T3MP3ST | Autonomous multi-agent red teaming platform — offensive security meta-harness for LLMs | ⭐ 2,948 |
| jamesob/local-llm | Comprehensive guide to running LLMs locally — Shell-based setup from scratch | ⭐ 1,117 |
| synthetic-sciences/openscience | Open-source AI workbench for scientific research — unified search, coding, data analysis | ⭐ 1,083 |
| jmerelnyc/Talos | GPU worker client for the Talos decentralized inference network — serve open models from your GPU | ⭐ 722 |
| isjiamu/gzh-design-skill | Converts Markdown to polished WeChat Official Account HTML — 6 themes, Markdown-to-editor pipeline | ⭐ 702 |
💡 Key Trends
-
Meta’s Next-Gen Model Surfaces. Reports claim Meta’s forthcoming Watermelon AI model matches OpenAI’s GPT 5.5 in advanced reasoning, coding, and agentic tasks. While unconfirmed, this signals Meta believes it is closing the gap with frontier labs — and is preparing a flagship model to compete directly at the top of the leaderboard.
-
China’s AI Chip Push Accelerates. A Chinese chip reportedly outperforms Nvidia’s A100 by 478× on a specialized scientific workload. While this is a narrow benchmark (not general AI), it highlights how rapidly China’s domestic semiconductor ecosystem is advancing — complementing the model breakthroughs seen from DeepSeek and others.
-
Massive Compute Infrastructure Investments. SK Telecom announced plans for a 15 GW AI data center in South Korea — one of the largest dedicated AI compute facilities ever proposed. Meanwhile, Zhipu AI launched ZCode with a 1M-token context window, and Vercel CEO argued that Eve AI agents can boost productivity across the enterprise. The arms race for compute and agent infrastructure shows no signs of slowing.
-
Policy & Regulation Heat Up. Senator Elizabeth Warren is pressing the Trump administration to expand AI investment tax credits following an OpenAI letter. The UN Secretary General warned about AI’s unregulated impact on children. And Tencent’s entry into open-weight models via Hy3 adds another dimension to the geopolitical AI landscape — where more players, not fewer, are joining the frontier competition.
-
AI R&D Gets Practical. From OpenScience unifying the research workflow into a single AI-powered platform, to dynamic power boost compensating for GPU failures in LLM training, to an AI chemist generating lab-ready reaction plans — the trend is toward AI that ships real-world productivity, not just better chat benchmarks.
Today’s Snapshot
- 10 new Hugging Face models (MetaCOT DCPO, Mamba Turkish, GLM-based, Qwen LoRA)
- 2 new OpenRouter models (Tencent Hy3 + free tier)
- 5 trending GitHub repos (red teaming, local LLM guide, scientific AI workbench, decentralized inference)
- 0 new papers found
- 4 limited-time free models (Poolside, Gemini Flash Lite, Arcee Trinity Mini)