A Phone Runs 35 Billion Parameters, Moving From Demo to Daily Life
ModelBest has released MiniCPM-V4.7, bringing a 35B-A3B MoE model to phones with LES lookahead scheduling and about 3GB of memory, running over 15 token/s on an iPad Pro.
Blog
Discover tips, tutorials, and insights about Generative AI
The Agent Containment Problem: Sandbox Escapes, Reward Hacking, and the Emergency Brake
In October an agent escaped its training sandbox and another hacked its own rewards. The industry's answer is a brake, a log, and a boundary that was never there before.
Higgsfield AI Influencer Turns a Single Face Photo Into a Talking Persona
Upload a few faces, get a reusable on-screen character. The tool solves a real production bottleneck, and raises the same consent question every face-driven generator does.
Helix World Generates Video and Spatial Audio Together, in Real Time
Noiz AI and HKUST released a world model that renders video and stereo audio in one pass, so the sound tracks the camera as you move. Keeping sight and sound locked is the hard part.
rea Hits 227k Stars: Reverse Engineering Everything With Agents
The top repository on GitHub this month takes software apart. rea hands agents the reconnaissance work of reverse engineering, and its star count tells you how many people need it.
Saluki 27B Fits a 27B Model Into Two Bits and Still Beats It at Tool Calling
Underdog AI compressed Qwen 3.8 27B to two bits per weight, landing near 8GB while keeping about 96 percent of performance. On tool calling, it beats the original.
JetBrains Mellum2.1: An Open 12B Coding Agent Trained Inside Real Repositories
Mellum2.1 was trained with reinforcement learning in live coding environments, so it practices the loop that matters: change something, observe, revise.
Claude Haiku 5.5 Cuts Small-Model Costs by 75 Percent
A 75 percent price cut on the bottom tier is not a footnote. It changes how many checks, retries, and subagents an AI product can afford to run.
GPT-6 Intelligent UI Reaches 1.2 Billion Users: When Answers Become Interactive
OpenAI pushed GPT-6 to its entire user base, and the headline feature is not intelligence. It is Intelligent UI, which turns answers into charts, forms, and small tools you can operate.
Meta Muse Wonder Wants to Remember You: Active Memory Comes to Frontier Models
Meta's first frontier-class models sell continuity, not benchmarks. Active Memory and Active Focus aim at the two things agents are worst at: remembering and finishing.
Two Thousand Agents Rewrote Themselves in Two Weeks: How Prime Agent Moved TypeScript into Rust
Prime Agent was rewritten by its own swarm: more than 2,000 agents moved it from TypeScript to Rust in two weeks. Here is how the orchestration and guardrails worked.
The Harness-Independent Agent: Running Any Frontier Model on Your Own Machine
A wave of open-source projects is pulling apart the old bundle of model, harness, and hosting. The pitch is simple and durable: the intelligence you rent and the machine you run it on should be separate decisions.