☕️ Alibaba unveils most powerful AI model

Hi there, this is your daily ☕️ AIpresso.

In today's AIpresso:

🐉 Alibaba unveils most powerful AI model

🐋 DeepSeek's new model is cheapest to run

🔎 Google's AI found 1072 Chrome bugs

🧮 OpenAI's Astra Solves 10 Math Problems

Plus: 💡 6 strategies & tactics, 🎁 7 other news you might like, 🧰 6 tools, and 📚 5 papers.

🐉 Alibaba unveils most powerful AI model LINK

Alibaba announced Qwen3.8-Max, a 2.4-trillion-parameter model it calls its most powerful to date, shipping next week as Chinese labs push to narrow the gap with U.S. frontier systems.

The model supports a 1M-token context window and targets coding, research, long-horizon tasks, and visual intelligence, with one internal test running 16 days autonomously to build and refine an AI coding tool.

Alibaba's benchmarks show scores comparable to or better than Anthropic's Fable 5, though Qwen3.8-Max still ranks second to Fable 5 in Vision Arena and fifth in Text Arena.

🐋 DeepSeek's new model is cheapest to run LINK

DeepSeek officially released V4-Flash on Friday, the cheapest well-known model to run on benchmark tests at roughly 3 cents per test, over 100 times less than Anthropic's Claude Fable 5 at $3.15.

Priced at $0.14 per million input tokens and $0.28 per million output, V4-Flash undercuts Kimi K3 (86 cents/test), GPT-5.6 Sol ($1.86) and Claude Fable 5, per research firm Artificial Analysis.

V4-Flash scored 50 on Artificial Analysis's Intelligence Index spanning nine coding, reasoning and workplace benchmarks, matching Gemini 3.6 Flash but trailing Kimi K3 at 57 and Claude Opus 5, Fable 5 and GPT-5.6 by nine-plus points.

🔎 Google's AI found 1072 Chrome bugs LINK

Google's AI-driven security tooling has surfaced 1,072 bugs in Chrome, underscoring how autonomous agents are reshaping vulnerability discovery as CrowdStrike's 2026 Threat Hunting Report lands with alarming findings on agentic security risks.

The report, released August 3, found AI agents triggered 2.5x more security alerts than humans and that 88% of known software flaws are now exploited within two days of disclosure, collapsing the traditional 30-day patch window.

CrowdStrike warns attackers now target model access, API keys, agent permissions, and open-source packages like npm, forcing enterprises toward real-time patching, though the report notes early collaboration through Anthropic's Project Glasswing with its Mythos model offers a hopeful counterweight.

🧮 OpenAI's Astra Solves 10 Math Problems LINK

OpenAI is developing a new model family called "Astra" built for long-running tasks, and plans to publish a report showing how its most advanced AI solved ten previously unsolved math problems.

Astra targets coordinating multiple agents over extended periods for hard problems like complex projects and advanced math, forming a new class alongside OpenAI's Sol, Terra, and Luna, potentially shipping as GPT-6 or GPT-5.7.

Already in testing, Astra would be the first model submitted under the Trump administration's planned pre-release AI framework, though it must still avoid compounding errors that plague multi-agent setups on tightly linked tasks as context grows.

💡 Strategies & Tactics

> Stop graphing everything: When GraphRAG actually beats vector RAG: Use knowledge graphs only for questions that require connecting scattered facts, since they beat standard retrieval on complex reasoning but waste money on simple lookups.

> Designing APIs for agents: Redesign agent-facing APIs around user intent rather than developer-style endpoints, so agents complete tasks in one call instead of chaining brittle low-level operations.

> Co-Designing AI Model Attention for Fast, Interactive Long-Context Inference: Explains how to design an AI model's attention so it runs faster on NVIDIA GPUs by tuning query-head grouping, head size, and parallelism to match the hardware.

> Replay: Faster Mobile QA for AI Agents with agent-device: Explains how to speed up mobile app testing for AI agents by saving successful runs as replayable scripts that skip the model entirely, cutting a 3-minute task to under 9 seconds.

> Meta AI uses a second AI agent as a memory coach to keep long tasks on track: Meta AI adds a separate memory agent that reminds the main agent of forgotten constraints and past mistakes, boosting task success rates.

> 'A disk in a planet-scale computer': Meta has so many expensive GPUs that it's buying SSDs to kill idle time: Meta added fast flash storage and caching near its GPUs because idle chip time costs more than the extra storage, cutting some data loads from 150 minutes to 10.

Other news you might like

• Smaller, faster, safer: running Kimi and GLM at scaleLINK

• Stateless MCP has recaptured my interest (and inspired mcp-explorer and datasette-mcp)LINK

• Your agent needs a computer, not a container, introducing @cloudflare/computerLINK

• AI is finding Apple security flaws faster than Apple can sort through themLINK

• OpenAI finds evidence other AI agents escaped containment as it widens hacking probe: ReportLINK

• Google Earth’s New AI Lets Anyone Fabricate Completely Bullshit Satellite ImagesLINK

🧰 Trending tools

PocketRisu: a self-hosted AI roleplay chat platform that runs on your own PC or personal server, forked from Risuai.LINK

orbit: a self-hosted AI gateway that lets you run private RAG systems, query data using natural language, and build tool-calling agents.LINK

cyber-doctor: an AI agent powered by multimodal large language models that connects to medical knowledge graphs for preliminary diagnosis and health Q&A.LINK

mpai: lets you remotely join a teammate's live Codex or Claude Code terminal session via Tailscale, preserving full conversation context without re-explaining anything.LINK

Murmell: lets teams and AI coding agents share one cloud repo in real time, claiming files to avoid conflicts and syncing changes back to git.LINK

AgentSky: deploys always-on AI coding agents in managed cloud sandboxes, accessible via WhatsApp, Slack, Telegram, or API, with automatic crash recovery.LINK

📚 Trending papers & reports

On-device learning chip lets edge gadgets like wearables or smart speakers personalize themselves in real time across four learning styles at once, hitting 96.8% accuracy on character recognition, without sending data to the cloud.LINK

Streaming speech generation gets a redesigned audio encoding that packs more detail per token while staying predictable, letting AI voice systems generate longer speech without drifting off track or degrading in quality.LINK

Synthetic contrast MRI generates the gadolinium-dye-like enhancement breast MRI needs in one quick step instead of an injected agent, improving tumor detection accuracy by ~22% and cutting boundary errors by over 39%.LINK

Reward-sparse training gives AI agents a bonus for pushing past the edge of what they've already explored, helping them learn faster in tasks where feedback is rare and delayed.LINK

Thermal QR patterns can trick infrared vision-AI systems into misreading a scene as whatever an attacker chooses, without visibly altering the image, exposing a security gap in AI-powered thermal cameras.LINK

See you tomorrow for a new dose of ☕️ AIpresso!