Daily AI News - 2026-06-29

🧠 DeepReinforce Open-Sources Ornith-1.0 Agentic Coding Models

DeepReinforce released Ornith-1.0 on June 26 — a family of open-source, self-improving coding models under the MIT license. The lineup spans four sizes: 9B Dense, 31B Dense, 35B MoE, and 397B MoE, all built on pretrained Gemma 4 and Qwen 3.5 foundations.

What sets Ornith apart: the model learns to generate its own RL scaffold alongside the solution. During training, it proposes a task-specific scaffold, then generates a solution conditioned on it — rewards flow back to improve both. The 397B flagship scores 82.4% on SWE-Bench Verified and 77.5 on Terminal-Bench 2.1, surpassing Claude Opus 4.7 and matching top open-source peers.

Three security layers guard against reward hacking: a fixed trust boundary, a deterministic monitor, and a frozen LLM judge. All variants ship with FP8 and GGUF quantizations for efficient deployment. The models are compatible with OpenHands, Hermes Agent, and OpenClaw.

🏢 Google DeepMind Talent Exodus Accelerates

Six senior researchers have left Google DeepMind for rivals since February, the sharpest talent loss streak in years. The most consequential departure is Denny Zhou, who founded Google Brain’s reasoning research group and spent eight years building multi-step thinking foundations — now at Meta Superintelligence Lab. Nobel laureate John Jumper (AlphaFold) joined Anthropic on June 19, followed by Jonas Adler and Alexander Pritzel on June 24. Dawn Song, the UC Berkeley AI security pioneer, joined Meta Superintelligence Lab as VP of AI Research on June 25, bringing her Virtue AI co-founders along.

The departures signal a structural shift: Google DeepMind’s pivot away from coding-focused research toward scientific AI has created a talent opening that Meta, OpenAI, and Anthropic are aggressively filling.

🤖 Google Integrates Computer Use Natively Into Gemini 3.5 Flash

Google announced native Computer Use integration into Gemini 3.5 Flash via API and the Enterprise Agent Platform — allowing models to directly control GUI elements, click buttons, fill forms, and navigate web interfaces. The feature ships alongside the broader Gemini Enterprise Agent Platform with native plug-ins from Adobe, Atlassian, Salesforce, ServiceNow, and Workday. Google also confirmed its AI coding strike team scope has expanded to “midtraining” — a response to talent losses and competitive pressure from Anthropic.

📈 Anthropic Files Confidential IPO, Approaches $1T Valuation

Anthropic has filed confidentially for an IPO on the US stock market, continuing its rapid ascent. The company’s run-rate revenue has surpassed $30 billion (up from ~$9B at end of 2025), with over 1,000 enterprise customers each spending over $1M annually on an annualized basis. The IPO filing follows a $65B funding round that valued the company at $965B post-money. On Polymarket, traders give Anthropic an 82% implied probability of having the best AI model by end of June 2026.

🔗 OpenRouter Launches Fusion API, Subagent, and Advisor Tools

OpenRouter unveiled three new server tools in June:

• Fusion API — Routes prompts to 3-5 models in parallel, synthesizing the best answer. Claims to match Claude Fable 5 at roughly half the cost. Outscored GPT-5.5 and Claude Opus 4.8 on 100 research tasks.
• Subagent — Lets a frontier model hand off self-contained tasks (summarization, data extraction, reformatting) to a smaller, cheaper worker model mid-generation.
• Advisor — Gives any cheap model the ability to consult a stronger one mid-response (e.g., GPT-4o Mini for routine work, Claude Fable when it matters).

🤗 Trending on Hugging Face

• sundaycoil/circuit-breaker — Experimental model exploring circuit-breaking techniques in neural networks. 2 likes.
• bijaykumarsingh/whisper-medium-commonvoice17-bn — Whisper medium fine-tune on CommonVoice 17 Bengali speech data. 2 likes.
• Miiche/visualrl-lora-m2-only — Visual reinforcement learning LoRA adapter for image-based RL tasks. safetensors format.
• Miiche/visualrl-lora-grpo — Companion Visual RL LoRA using GRPO optimization, safetensors format.
• hdhruva/act_cleandesk_v2_20260628 — Robotics ACT policy for clean-desk manipulation tasks, built with lerobot. safetensors format.
• jin8191/gemma-2b-brain-v3 — Gemma 4 based image-text-to-text model fine-tuned for neuroscience-style reasoning. Published with unsloth.
• peaceinlove/long-memory-connectaiLecture-season1-v2 — Long-context model fine-tune on ConnectAI lecture content, built on Gemma 4 with unsloth.
• Bridoxd/Brido_V7_RTRL — Real-Time Recurrent Learning (RTRL) experiment, safetensors format.
• Ba2han/experimental4 — Mistral-based experimental text-generation fine-tune, safetensors format.
• shelock/watermark-trustmark-model — Watermark detection and trustmark verification model for AI content provenance.

⭐ GitHub Trending: AI Edition

• winsznx/theeleven — Eleven autonomous AI agents running live football prop markets on X Layer via a custom Uniswap v4 hook. DeFi meets AI agents meets sports betting. ⭐ 702 — TypeScript
• benchflow-ai/awesome-evals — A curated library of best resources for building and evaluating AI agents: papers, blogs, and frameworks. ⭐ 565
• Pluviobyte/video-production-skills — Reusable AI video production skills library for creation, recreation, motion design, and QA workflows. ⭐ 366 — Python
• tutusagi/ai-fishing-game — 🎣 Deterministic text-based fishing mini-game for AI companions. Single file, zero dependencies. ⭐ 274 — Python
• eli-labz/Godcoder — A local-first, open-source coding agent for your desktop. Bring your own LLM key; your code stays on your machine. ⭐ 245 — Rust

💡 Key Trends

1. Agentic Coding Models Go Open-Source. DeepReinforce’s Ornith-1.0 release marks a new wave of self-improving coding agents. The RL-over-scaffold approach — where models learn to optimize their own tool-use harness — could reshape how agent frameworks are built. MIT licensing and full open weights lower the barrier for teams to deploy production-grade coding agents.

2. AI Talent War Intensifies. Six DeepMind researchers lost to Meta/OpenAI/Anthropic in five months, on top of Meta’s earlier poaching of seven OpenAI Zurich researchers. The talent market is being driven by a structural shift: labs that focus on coding agents vs. scientific AI are competing for the same small pool of senior researchers.

3. Anthropic’s Ascent Continues. From a $65B funding round at $965B valuation to confidential IPO filing, $30B revenue run-rate, 1,000+ million-dollar enterprise customers, and Polymarket’s 82% confidence in best-model status — Anthropic is having a defining quarter.

4. Google Responds With Product, Not Poaching. Rather than matching the talent war head-on, Google is shipping: native Computer Use in Gemini 3.5 Flash, expanded agent platform with enterprise plug-ins, and an expanded “midtraining” coding strike team. The strategy: make the platform the moat, not the people.

5. OpenRouter’s Tool Ecosystem Matures. With Fusion (multi-model synthesis), Subagent (model delegation), and Advisor (model consult), OpenRouter is building the middleware layer for agentic AI — letting developers compose models in increasingly sophisticated patterns without managing infrastructure.

Today’s Snapshot

• Trending HF Models: 10
• New OpenRouter Models: 0 (3 new server tools)
• Trending GitHub Repos: 5
• Total Items: 15
• News Highlights: DeepReinforce Ornith-1.0, Google DeepMind talent exodus, Anthropic IPO filing, Google Computer Use in Gemini 3.5 Flash, OpenRouter Fusion/Subagent/Advisor tools