Google reportedly cut off Meta's access to Gemini models. HP signed a strategic AI deal with OpenAI. And Alibaba's Qwen team published a technical report on their new image reasoning model.
Google has reportedly limited Meta's ability to use its Gemini AI models, according to a Financial Times report picked up by CNBC. That's a significant move — it signals that frontier model providers are starting to draw competitive lines around who gets API access, and at what scale. If you're building on top of any single provider's models, this is a reminder that access is not guaranteed.
Google limiting Meta's Gemini access is the most concrete business story today. The FT report (cited by CNBC) doesn't give a full technical reason, but the implication is competitive: Google doesn't want Meta training on or deeply integrating Gemini when Meta runs its own frontier model program (Llama). For indie builders, the practical lesson is diversification — don't build a product that only works on one provider's models.
HP Inc. announced a strategic partnership with OpenAI called Frontier. The announcement is live on OpenAI's site. Details are thin — no pricing, no specific product launches named — but it positions HP hardware and enterprise services alongside OpenAI's models. This is the kind of deal that matters if you sell into enterprise IT, less so if you're building consumer tools.
Alibaba's Qwen team published the Qwen-Image-2.0-RL Technical Report on Hugging Face. This covers a reinforcement-learning-trained vision model — the RL part means the model was trained with reward signals rather than pure supervised fine-tuning, which tends to improve reasoning on visual tasks. The paper is out; the model weights are not yet confirmed as publicly downloadable. Worth watching if you build anything involving image understanding or document parsing.
Also on Hugging Face: deepreinforce-ai posted Ornith-1.0-397B-FP8, a 397-billion-parameter model in FP8 format. It's ranked #49 on the models leaderboard. There's no accompanying paper or blog post linked, so provenance is unclear. At 397B FP8 you'd need serious hardware to run it locally.
llama.cpp hit build b9837 this week. If you run local models on CPU or Apple Silicon, keeping llama.cpp updated is just good hygiene — each build tends to include quantization fixes and speed improvements. Cline (the VS Code AI coding agent) released cli-v3.0.32. LobeHub hit v2.2.9. None of these are major version bumps, but if you use any of them daily, update.
On Hacker News, the EU's open-source release of Ten-Year Network Development Planning (TYNDP) tools got 158 points. The repo is at open-energy-transition/open-tyndp on GitHub. Not directly an AI story, but if you're building in the energy or infrastructure space, these are real planning datasets and models that were previously locked inside EU institutions.
1. Test your model provider fallback. Pick one project that calls a single AI API and add a second provider as a fallback. The Google-Meta story is a good prompt to do this now, not after something breaks.
2. Try the Qwen-Image-2.0-RL paper's benchmark tasks on your own image data using whatever vision model you currently use (GPT-4o, Claude 3.5 Sonnet, Gemini 1.5 Pro). Note where it fails. When Qwen-Image-2.0-RL weights drop publicly, you'll have a real comparison baseline ready.
3. Update llama.cpp to b9837 and re-run your favorite local model. Check if inference speed changed on your hardware. Takes 10 minutes and often gives a free speed bump.
Confirmed: Google limiting Meta's Gemini access was reported by the Financial Times and picked up by CNBC. The CNBC article is the source we have.
Not independently verified by CBW: We have not seen the full FT report, and the specific technical or contractual reason for the restriction has not been disclosed publicly.
Confirmed: HP's Frontier partnership with OpenAI has a live announcement page on openai.com as of today.
Not independently verified by CBW: The Qwen-Image-2.0-RL model weights are not confirmed as publicly downloadable. The technical report is on Hugging Face Papers but we have not found a model card with downloadable weights.
Worth noting: Ornith-1.0-397B-FP8 from deepreinforce-ai has no linked paper or blog post. Its training data, license, and benchmarks are unverified by CBW. Treat with caution.
Source: CNBC — Google limits Meta's use of Gemini — https://www.cnbc.com/2026/06/28/google-limits-metas-use-of-its-gemini-ai-models-ft-reports.html
Source: OpenAI — HP Frontier partnership announcement — https://openai.com/index/hp-frontier-partnership
Source: Hugging Face Papers — Qwen-Image-2.0-RL Technical Report — https://huggingface.co/papers/2606.27608
Source: Hugging Face — Ornith-1.0-397B-FP8 model — https://huggingface.co/deepreinforce-ai/Ornith-1.0-397B-FP8
Source: GitHub — llama.cpp b9837 — https://github.com/ggml-org/llama.cpp
Source: GitHub — EU open-tyndp energy planning tools — https://github.com/open-energy-transition/open-tyndp
Get tomorrow's best AI project in your email. With a guide that works. Free. No spam.