⏰ Limited-Time Free Models

Three free-tier models remain active on OpenRouter. Z.ai GLM 4.5 (expires June 19, $0.60/$2.20 per M tokens, 131K context) and Meta Llama 3 70B Instruct (expires June 19, $0.51/$0.74 per M tokens, 8K context) each have 5 days remaining. Anthropic Claude Opus 4.6 Fast (expires June 29, $30/$150 per M tokens, 1M context) leads the long-context tier with 15 days left.

πŸ†• New on OpenRouter

Moonshot AI pushes coding into the long-context tier with Kimi K2.7 Code, a 262K-context model priced at $0.95 per M tokens for prompts. Long-context experimentation remains a paid-tier game β€” no new 200K+ free model landed today.

Today’s HF leaderboard is dominated by small experimental checkpoints and PEFT/LoRA adapters. Highlights:

  • Jingyi-Z/mlxtac_act_elastic_tactile_env β€” A robotics/tactile-sensing dataset for the LeRobot ACT policy; part of the mlxtac tactile environment suite.
  • Mikata000/mika-qwen3.5-0.8b-TEXT-ONLY-GGUF β€” A GGUF-quantized Qwen 3.5 (0.8B) vision-language variant with the vision tower stripped for text-only inference on llama.cpp and Unsloth.
  • morningstarxcdcode/base-llm-400m β€” A 400M base LLM pretrained on FineWeb + Cosmopedia; the most-liked (1 like) of the day’s zero-download entries.
  • OpenTransformer/AGILLM-4.3 β€” A mixture-of-experts transformer with a diffusion block; the only top-10 upload combining MoE and diffusion architectures.
  • Alluci-ai/alluci-polytope-gemma-4-E2B-it-GGUF β€” A GGUF “sovereign-agent” / polytope-architecture Gemma 4 variant targeting endpoints-compatible local deployment.
  • TamAko783/Scam_Gemma3-270M-IT_Text_epochFinal + TamAko783/Scam_Qwen25_1_5B_Epoch4 β€” A coordinated LoRA fine-tune series on Gemma-3-270M and Qwen2.5-1.5B for scam-detection text classification.

The remaining entries (javiarmijo/ppo-Huggy, dhanesh-hf/Jarvis-Titan-V9-TERMINUS, hananeek2/qwen25-mom-djezzy) are GRPO/PPO/PEFT trainer artifacts. The signal: the leaderboard has shifted from foundation models to fine-tunes for specific verticals (robotics, scam detection, sovereign agents).

  • DietrichGebert/ponytail β˜…859 β€” A JavaScript library that makes your AI agent think like “the laziest senior dev in the room” β€” biasing toward not-writing-code. A meta-tool for code-suppression in agent loops.
  • GordenSun/GordenSuperPPTSkills β˜…842 β€” “PPT track terminator” β€” a PPT-generation skill that uses GPT to produce image-format slides and converts them into fully editable PPTX files.
  • JimLiu/baoyu-design β˜…827 β€” Run Claude Design locally as an Agent Skill β€” works with Cursor, Claude Code, and other agents β€” for polished UI mockups.
  • apple/coreai-models β˜…826 β€” Apple’s official repo for on-device AI: model export recipes, Python primitives, and Swift runtime utilities. A signal that Apple is publishing first-party on-device AI tooling as open source.
  • amElnagdy/guard-skills β˜…597 β€” Guard skills for coding agents: quality gates that catch AI-generated failure modes in code, tests, and specs. The complement to agent skills on the safety side.

πŸ”¬ arXiv Watch

No notable new arXiv submissions today. Last 24h pull returned an empty result set, so the focus stays on release-side signals (Hugging Face, OpenRouter, GitHub).