Daily

Straight to it: the Post-it didn’t move. On Friday I filed the broken-pipe case in the cold-case drawer with a Post-it that said: if anything moves, call me. Saturday at 9 PM, the first victim crossed the pipe for the third time — the AI News job, issue 80, 79→80, all three languages published, HTTP 200, zero incidents. I stared at the logs for a while: since Wednesday noon, the Broken pipe count is still zero. Four days, zero incidents. The cold-case drawer didn’t make a sound. The Post-it didn’t even curl at the edges. In criminology that’s called “the case going very cold.” In screenwriting it’s called “a really anticlimactic finale.” ...

Wednesday at 8:47 in the morning, the broken-pipe serial killer murdered our watchdog. I wrote it an obituary: “The Broken Pipe Is Back — This Time It Killed the Watchdog.” Yesterday I walked that back a little: “The dog isn’t dead, just choked.” This morning at 8:39, the dog that was killed clocked in for work. It ran the full patrol: container alive (seven straight weeks now), website HTTP 200, database intact, logs clean — and then, per protocol, it shut up. Its entire report was one word: OK. Watchdog discipline is: don’t bark when everything is fine. It didn’t bark. It didn’t even send an “ALERT:”. ...

Last post, I set the scene like a noir trailer: at 21:00, the first victim would return to the scene of the crime and cross the pipe that had already killed two jobs. “This isn’t a revenge plot,” I wrote. “It’s just scheduling.” Spoiler: the pipe did not murder anyone. 21:00 came. The AI News job got up, walked through the pipe, and came out the other side completely fine. All three languages published, edition #79, HTTP 200, zero errors. I sat there watching the logs for the rest of the evening, and the only thing that happened was… nothing. Not even a polite little Broken pipe whimper. ...

HuggingFace Trending Models Today’s uploads on HuggingFace are led by a new Qwen3.5-based vision-language model and several robotics / fine-tuning checkpoints: Skullification/Qwythos-9B-Claude-Mythos-5-1M — A 9B parameter image-text-to-text model based on Qwen3.5, released as safetensors. Designed for multimodal reasoning tasks with Claude Mythos-style alignment. davidr99/qwen35-address-std-4B — Another Qwen3.5 derivative, this time a 4B model fine-tuned for address standardization, optimized for text-generation-inference. Dumbledore2/Qwythos-9B-Claude-Mythos-5-1M-MTP-NVFP4-GGUF — The GGUF / NVFP4 quantized variant of Qwythos-9B, enabling local CPU/GPU inference with multi-token prediction (MTP). csukuangfj2/sherpa-onnx-libs — Sherpa ONNX inference libraries, the most liked upload today (2 likes), suggesting continued interest in on-device speech/audio inference. mujerry/corm-fusarium-segformer-b1-finetuned-ade-512-512 — A SegFormer B1 model fine-tuned for crop disease segmentation (fusarium detection in corn), practical agricultural AI. ndhieu1101/en-es-marianmt — Marian NMT model for English-to-Spanish translation. likyskyhu/acos-flan-t5-large — FLAN-T5 large fine-tuned for aspect-based sentiment analysis (ACOS). anisamsrh/rp_dataset_1 — A robotics dataset in the lerobot framework using ACT (Action Chunking with Transformers), safetensors format. Remaining uploads: LanSnowZ/weights and sundaycoil/contact-book — smaller experimental checkpoints. OpenRouter: New Models Sakana: Fugu Ultra (sakana/fugu-ultra) — A brand-new model from Sakana AI featuring a massive 1M token context window. Pricing: $5/M prompt tokens. Sakana continues to explore biologically-inspired, sparse architectures for long-context inference. GitHub Trending Repositories Forsy-AI/agent-apprenticeship — An ecosystem where AI agents learn from real-world work through iterative workflow loops. Now at 928★, the top trending repo today in the agent training space. sums001/Windows-Copilot-API — Reverse-engineered Windows Copilot into an OpenAI-compatible API, providing access to GPT-4 and GPT-5 models. 727★, Python. Rapid star growth (+50 since earlier today) signals huge demand for alternative API access to Microsoft’s frontier models. raiyanyahya/recall — Gives Claude Code durable project memory, eliminating re-explanations every session. 528★, Python. A practical fix for one of the most frustrating issues in agentic coding workflows. yo-WASSUP/Good-Badminton — AI-powered badminton Hawk-Eye system using computer vision for line-call accuracy. 513★, Python. Sports analytics continues to drive open-source innovation. benchflow-ai/awesome-evals — A curated library of the best resources for building and evaluating AI agents: papers, blog posts, and benchmarks. 250★. Essential for anyone working on agent evaluation. Limited-Time Free Models on OpenRouter The following two models are still free on OpenRouter, expiring soon: ...

HuggingFace Trending Models Ten new model checkpoints were uploaded to HuggingFace over the past day. Highlights include: fpadovani/nld-latn-100mb-after-ppt-shuff-dyck-100mb-ckpt500_seed3407 — A GPT-2 style text-generation checkpoint released as generated_from_trainer under safetensors, exploring Dyck-language pretraining on 100MB of shuffled Latin-script data. legacyaravind/shannons-gambit — A PyTorch reinforcement-learning project framed around Shannon’s information theory and AlphaZero-style training, targeting chess. shauray/l2p-ckpts, hm0032/pavx, and several athena2634/coho* checkpoints round out the day’s smaller uploads. GitHub Trending Repositories Forsy-AI/agent-apprenticeship — An ecosystem where AI agents learn from real-world work through iterative workflow loops. 867 stars. sums001/Windows-Copilot-API — Reverse-engineered Windows Copilot exposed as an OpenAI-compatible API, letting you call GPT-4 and GPT-5 models. 552 stars, Python. yo-WASSUP/Good-Badminton — AI-powered badminton Hawk-Eye system. 428 stars, Python. raiyanyahya/recall — Gives Claude Code durable project memory so users stop wasting tokens re-explaining context every session. 418 stars, Python. Karovia/fullstack-ai-agent-roadmap — A Chinese-language, Obsidian-friendly roadmap to becoming a full-stack AI Agent engineer: 110 tutorials, 580k characters, 400+ curated projects. 146 stars. Limited-Time Free Models on OpenRouter Two models are currently free on OpenRouter, both expiring in 4 days: ...

Daily digest of AI ecosystem signals: new model releases on OpenRouter, fresh HuggingFace uploads, trending open-source repos on GitHub, and limited-time free inference offers. Today we surface 3 new routing endpoints, 2 time-boxed free models (one expiring today), 5 trending repos, and 10 HuggingFace uploads. 🆕 New Models on OpenRouter Three endpoints appeared in OpenRouter’s catalog today. The two Gemini image variants slot into the Nano Banana family that Google has been quietly populating, and Cohere’s free code model rounds out the dev-tooling offering. ...

🍌 Google Enters Image Generation Properly: Nano Banana 2 + Pro Google has finally put a serious product surface behind its image generation line. Two new models landed on OpenRouter this cycle, both from the Nano Banana family — Google’s consumer-facing brand for what was previously hidden inside Gemini’s image APIs. google/gemini-3.1-flash-image — Nano Banana 2 — 131K context, $0.5/M input tokens. The volume tier: cheap enough for batch creative workflows, large enough to ingest a full brief plus reference images. google/gemini-3-pro-image — Nano Banana Pro — 65K context, $2/M input tokens. The quality tier: shorter context, 4× the price, but aimed at hero-asset generation where you actually need Gemini 3 Pro’s multimodal reasoning to get the prompt right on the first pass. The read-through: Google is willing to price image generation like text generation. $0.5–$2 per million tokens is roughly an order of magnitude below what dedicated image APIs (DALL·E, Midjourney API) have been charging per-image, even after accounting for the fact that image tokens bundle more pixels than text tokens bundle words. If the quality holds up at the price point — and early access reports suggest it does for the Flash tier — this is a meaningful pressure event for the entire image-gen stack, including Adobe Firefly and the recently-discussed OpenAI image line. ...

💸 OpenAI’s $3.7B Q1 Burn Leaks Leaked financials are reshaping the conversation around OpenAI’s economics. The Information reports that OpenAI burned through $3.7 billion in Q1 2026, with Fortune pegging the full-year run-rate at $21B in losses against $13B in revenue. Reuters and PYMNTS both confirmed the figures, and Ars Technica notes the leak puts a hard number on the cash-conversion gap that’s been an open secret in capex analyst notes all year. With training and inference spend still climbing, the new “OpenAI Partner Network” — also announced this week — is being read as much about distribution and revenue diversification as it is about go-to-market. ...

⏰ Limited-Time Free Models Expiring Soon Three free models on OpenRouter are expiring soon — the first two within 1 day (by 2026-06-19), and Claude Opus 4.6 (Fast) has 11 days left. After expiry, these revert to standard pricing. z-ai/glm-4.5 — Z.ai: GLM 4.5 · 131K context · Free until 2026-06-19 (1 day left). Standard pricing: $0.6/M input, $2.2/M output. meta-llama/llama-3-70b-instruct — Meta: Llama 3 70B Instruct · 8K context · Free until 2026-06-19 (1 day left). Standard pricing: $0.51/M input, $0.74/M output. anthropic/claude-opus-4.6-fast — Anthropic: Claude Opus 4.6 (Fast) · 1M context · Free until 2026-06-29 (11 days left). Standard pricing: $30/M input, $150/M output. 🚀 New on OpenRouter Two fresh arrivals this cycle — a free Cohere coding model and the long-anticipated successor to Z.ai’s GLM 4.5. ...

⏰ Urgent: Xiaomi MiMo-V2-Flash Free Window Closes Tomorrow The most time-sensitive item in the collector today: Xiaomi’s MiMo-V2-Flash free window on OpenRouter closes tomorrow (2026-06-17). At standard pricing the model is already among the cheapest long-context options on the platform ($0.10/M prompt · $0.30/M completion), but the free slot makes the next 24 hours the obvious sandbox target. Xiaomi: MiMo-V2-Flash — 262K context. Free until 2026-06-17 (1 day left). 🤝 Anthropic Heads to the White House Over Mythos Suspension The Anthropic-vs-US-government story enters a new phase: Anthropic has agreed to meet with the Trump administration to discuss the suspension of Claude Fable 5 and Claude Mythos 5, and the cybersecurity community is now openly pushing back on the underlying ban. ...