
MiniCPM-V 4.7: A 35B MoE Vision Model That Runs on an iPhone
OpenBMB MiniCPM-V 4.7 35B-A3B is an Apache-2.0 vision MoE with 13+ tok/s on iPhone 17 Pro Max. What is verified, how to run it, and caveats.

Expert profile
Founder & AI Product Leader
Yash Thakker is a Generative AI expert with over 12 years of experience in product leadership and technical strategy. As the founder of explainx.ai, he has taught over 300,000 learners and built AI platforms serving millions of users globally. He specializes in Agentic AI, Multimodal RAG, and the intersection of LLMs with consumer hardware.

OpenBMB MiniCPM-V 4.7 35B-A3B is an Apache-2.0 vision MoE with 13+ tok/s on iPhone 17 Pro Max. What is verified, how to run it, and caveats.

At OCP 2026 NVIDIA opened Vera Rubin rack designs via MGX 1.2 (250kW), Spectrum-6 102.4T switches, USFP optics and a grid resilience spec.

ReplaySSM from NVIDIA and Tri Dao caches SSM inputs instead of state: up to 1.8x kernel speedups and 2.1x speculative throughput on hybrid models.

Ondas and Palantir formed Regna AI to build fine-tuned open-weight models that governments own and run on drones at the edge. What is announced.

OpenCode 2.0 adds Code Mode, background subagents, 80+ internal plugins and a new SDK on one shared server. What breaks and how to upgrade safely.

Palantir CTO Shyam Sankar called the OpenAI Hugging Face hack a configuration error and a defective product. What OpenAI and METR documented instead.

QGIS merged QEP 428 on Oct 7, 2026: no AI contributions. Why maintainers did it, the scope questions left open, and what other projects can learn.

Replica is an MIT set of 11 Claude skills that reverse-engineer, rebuild, test and rebrand an app. Install steps, workflow and legal limits.

Subquadratic launched SubQ Snap, a decision model that scores fixed answers over 384K tokens in under a second for 100K inputs, at $0.05 per million.

text-to-cad is a free MIT plugin that lets Claude Code, Codex, Cursor, Gemini and Grok build STEP, STL and 3MF models and run DFM checks.

Linus Torvalds says Claude Code wrote the code and tests for a page-cache patch series. What the branch changes, the reported numbers, and what is unverified.

CBC found WE NEWS runs 121 AI correspondents, three posing as northern Indigenous reporters. What is confirmed, what is not, and how to spot them.