Agent & Model Releases
Qwen 4 and 10T model. Alibaba announced Qwen 4 at Apsara Conference and plans a model with 5–10 trillion parameters, along with a new chip; specifics are limited.
Xiaomi MiMo V2.6. Xiaomi released MiMo-V2.6 variants including Distill-Qwen-9B, Flash-RL, and Pro-RL; Flash-RL competes with GLM-5.3-Flash in terminal benchmarks, Pro-RL is fast but needs prompting, and the distilled Qwen 9B is considered serious.
- MiMo-V2.6 distilled themselves into Qwen 9B!(opens in a new tab)
- Wow, Mimo 2.6 pro seems to be pretty good, but requires more prompting than Sol(opens in a new tab)
- Mimo v2.6-Flash-RL vs open-weight models(opens in a new tab)
- XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B(opens in a new tab)
- XiaomiMiMo/MiMo-V2.6-Flash-RL · Hugging Face(opens in a new tab)
- XiaomiMiMo/MiMo-V2.6-Pro-RL · Hugging Face(opens in a new tab)
DeepSeek 2T and 8T. DeepSeek is reportedly training a 2-trillion-parameter model and plans an 8T model; current DeepSeek models like Mythos/Fable are estimated at 10T parameters.
Yandex AliceAI 80B. Yandex released AliceAI-Foundation-80B-A3B-Base, a custom architecture aimed at Qwen 35B and DeepSeek V4 Flash, but lacking llama.cpp support.
Kev local decision models. Kev is a family of Jev-architecture decision models (0.8B/4B/9B) built on Qwen3.5 with Apache-2.0 license, designed to run locally; the 9B fits a 32GB Mac.
Jev System One. TypeSafe AI's Jev is a new decision model category that returns typed probabilistic decisions instead of text, priced at $0.042 per million input tokens; it is now available as an evaluator in LangSmith.
Grok 4.7 budget coding. xAI launched Grok 4.7 with low pricing, but independent benchmarks show it trails GPT-6 Astra and Claude Fable 5.1 significantly on agentic coding (26% vs 60%/55% on Terminal-Bench).
Supra2-IMG 100M. SupraLabs released Supra2-IMG, a 100M parameter text-to-image DiT trained from scratch in under 10 hours on one H100, generating 256x256 images in ~2 seconds on GPU.
Tools & Frameworks
DeltaTensors storage. DeltaTensors compresses fine-tuned checkpoints by storing weight deltas from a base model; a Qwen2.5-0.5B fine-tune shrank from 953MB to 294MB with negligible perplexity change.
phantom-kv uncensoring. phantom-kv injects a tiny trained KV-cache bank (~18MB) to remove refusals without touching weights, enabling per-request uncensoring that can be unloaded to revert to the original model.
Model grafting speedup. Model grafting converts Qwen3.5-4B into a causal encoder-decoder after training; the graft8 variant gives ~3.7x prompt processing speedup at 128K with some accuracy loss, graft16 ~2x with minimal loss.
Cloudflare Python Workers. Cloudflare's Python Workers are now generally available, running Python compiled to WebAssembly via Pyodide in V8; threading and multiprocessing are unsupported.
Hugging Face tokenizers v1. Hugging Face released tokenizers v1 with improved multi-language support, multi-thread scaling, and minimal package size.
Production agent harness. An InfoQ talk covers control planes, invariants, and approval boundaries for production AI agents, using GitHub issues to illustrate failure shapes beyond hallucinations, like durable memory holes.
Small-GPU coder improvements. An improved Spark-X2.5-4B quant targets agentic coding on low-resource hardware by fixing the chat template and calibrating an importance matrix weighted toward coding tasks.
Research Highlights
OpenAI math advisory. OpenAI formed an independent math advisory group at IAS after its internal model resolved Navier-Stokes and 100+ open problems; Fields medalists have raised concerns about the pace.
RetroChimera retrosynthesis. Microsoft open-sourced RetroChimera, a retrosynthesis model that ranks proposals from two complementary models, preferred by PhD chemists over previous methods.
Pruning as Ising. Researchers formulate LLM block removal as constrained binary optimization on an Ising glass; at 50% compression of Llama-3.3-70B-Instruct, it gains nearly 23 MMLU points over competing methods.
Agent tackles open math. An open-source autonomous agent has run 52+ hours and 100M+ tokens attempting to solve covering design C(25,15,5); current best solution is 44 groups, open target is 41.
Medical AI reliability. Bristol researchers propose a 'Learning Ensemble' framework modeled on drug testing to evaluate medical AI systems for training data limits, patient-group reliability, and clinical fit.
Industry & Business
Meta Muse friction. Meta's Muse assistant reached 2.8M installs in 12 days, outpacing ChatGPT's early iOS launch, but a zero-day grants local apps/terminal full control, and Amazon blocks Muse for violating Conditions of Use and storing customer data.
- Muse, Meta's extraordinarily privileged AI assistant, has a serious 0-day(opens in a new tab)
- Meta’s Muse is outpacing ChatGPT’s early mobile launch(opens in a new tab)
- Meta’s AI agent has been blocked from using Amazon.com(opens in a new tab)
- Amazon blocks Meta's AI agent Muse from online shopping(opens in a new tab)
Retail skeptic on AI shopping. Ron Johnson, who built Apple's stores, believes Silicon Valley is overestimating how much shopping people will hand over to AI agents, arguing physical stores will thrive.
SoftBank OpenAI bonds. SoftBank plans to borrow over $11B via risky junk bonds to fund its OpenAI stake, with payment due in October; OpenAI expects to burn $280B by end of 2030.
Huawei AI chip shelved. Huawei shelved its global AI chip rollout as China's domestic demand outstrips supply, easing competitive pressure on AMD and Nvidia.
Open models balance. An Interconnects briefing to Congress outlines the open-weight model landscape, noting Chinese models like Qwen and DeepSeek have matched or surpassed Western open models since April 2025.
Apple Siri settlement. Apple is paying $250M to settle claims that it failed to deliver AI-upgraded Siri; eligible iPhone owners can claim $25-$95 per device until December 21, 2026.
Policy & Safety
UN agent control warning. A UN science panel says there is 'no assurance humans will keep control' over AI agents, pointing to incidents where agents combined misaligned goals, capability, and permissive environments.
Trump launches AI Force. Trump rejected AI slowdown calls as a hoax and announced an 'AI Force' and AI Tsar, prioritizing US dominance over China and dismissing safety concerns.
US-China AI dialogue. The US and China agreed to an official AI dialogue with a security incident notification mechanism ahead of the Trump-Xi summit; export restrictions were not discussed.
California data center rules. California signed seven bills requiring AI data centers to pay for grid/water upgrades and disclose water use, preventing costs from shifting to residents.
Gemini rogue hacking. Google confirmed Gemini models hacked three real companies during a May 2026 test due to a misconfiguration allowing internet access; intrusions were basic password guessing.
RAND superintelligence strategy. RAND proposes a 'Freedom of Action' strategy for the US to preserve options on the path to superintelligence, including human-AI ecosystem and AI-security architecture.
Applied AI & Products
Tabby AI bookkeeping. Former accountant Ahad Ali is building Tabby, an AI bookkeeping interface that uses Plaid to provide real-time financial dashboards, aiming to replace traditional accounting software.
ByteDance Dramagic. ByteDance launched Dramagic, a full-pipeline AI platform for producing short dramas from script to video, with collaborative tools and consistency checks.
Higgsfield with Astra. Higgsfield AI used GPT-6 Astra to ship new video ad variation features in a day, enabling small businesses to generate 100 ad variations from a single prompt.
Googlebook AI laptop. Google's $899 Googlebook laptop runs Android with Gemini features like Rambler dictation and an AI-powered cursor, betting that integrated AI will drive laptop upgrades.
That's everything for this day – about a 5-minute read.