New AI model releases and research
29 articles
Alibaba released HappyHorse 1.1 on June 23, pushing the model to second place on global video generation leaderboards. The timing is sharp: OpenAI's Sora has shut down and ByteDance's Seedance stalled internationally, leaving a wide gap for rivals to fill.
NVIDIA released Nemotron 3 Ultra, a 550B-parameter open-weights MoE model that ranks #1 among US open-weights models on Artificial Analysis's Intelligence Index, featuring a native 1M-token context window and full release of weights, training data, and recipes under a permissive commercial license.
Apple unveiled AFM 3 at WWDC 2026: five foundation models ranging from a 3B on-device model to a cloud Pro tier running on NVIDIA GPUs in Google Cloud, all while maintaining Private Cloud Compute privacy guarantees.
Jonas Adler and Alexander Pritzel — both AlphaFold veterans and Gemini contributors — are set to join Anthropic, Bloomberg reported June 24, marking the fourth and fifth senior departures from Google in a single week.
The European Commission selected the Domyn-led EUROPA consortium to develop a sovereign open-source AI model covering all 24 EU official languages at frontier scale — over 400 billion parameters.
Google's Gemini 3.5 Pro has slipped past its June 2026 general availability window, with release now expected in July. The flagship model was publicly promised for June by CEO Sundar Pichai at Google I/O.
OpenAI and Broadcom revealed Jalapeño, OpenAI's first custom ASIC built specifically for LLM inference. Developed in nine months with AI assistance, the chip targets large-scale deployment by late 2026 and is designed to reduce dependence on Nvidia.
Anthropic pulled back a planned billing change for Claude Agent SDK usage on June 15 — the same day it was scheduled to take effect — citing developer pushback and competitive dynamics ahead of its IPO.
Claude Fable 5 leaves included subscription tiers on June 23 and shifts to usage-based billing. Developers also face a new refusal response type and an always-on Adaptive Thinking mode that break assumptions built for older Claude models.
June 22 opens the primary Polymarket prediction window for GPT-5.6 at 83% probability. Leaked staging logs point to a 1.5-million-token context window and a redesigned reward audit pipeline targeting the alignment failures that plagued GPT-5.5.
xAI's Grok 4.3 is now generally available on Amazon Bedrock, adding a genuine third-party frontier alternative to GPT and Claude for AWS-native enterprise teams.
Prediction markets are pricing GPT-5.6 at 83% odds for a June 22–28 launch. OpenAI's chief scientist has reportedly confirmed the model is a meaningful step beyond GPT-5.5, with improvements in reasoning and agentic tasks.
Anthropic launched Claude Corps, a $150 million program placing 1,000 early-career AI fellows inside U.S. nonprofits. Fellows earn $85,000/year fully covered by Anthropic; the first 100-person cohort launches October 2026.
DeepSeek released V4 Pro in April preview — a 1.6 trillion total-parameter MoE model with a 1M-token context window, MIT license, and SWE-bench Verified score of 80.6%, priced at $0.87/M output tokens, roughly one-tenth of GPT-5.5.
Gemini 3.5 Pro, which Sundar Pichai promised at Google I/O would ship 'by next month,' remains in limited Vertex AI enterprise preview as of June 19 — with 11 days left in the window and prediction markets giving roughly 50–55% odds of a before-June-30 release.
Google has launched its first new smart speaker since 2020, placing Gemini AI at its core. The $99 device ships June 25 and targets multi-step conversational home control rather than simple voice commands.
Alibaba open-sourced Qwen3-Coder-Next, an ultra-sparse Mixture-of-Experts coding model with 80B total parameters but only 3B active at inference time, delivering up to 10× throughput versus dense equivalents and a 70.6% SWE-Bench score.
Andrej Karpathy, OpenAI co-founder and former head of Tesla AI, announced he is joining Anthropic to build and lead a new team using Claude itself to accelerate pre-training research.
NVIDIA unveiled Cosmos 3 at Computex 2026 — a fully open omni-model that handles text, image, video, sound, and action in a single system, compressing physical AI training cycles from months to days.
OpenAI released LifeSciBench, a 750-task benchmark built by 173 biotech and pharma scientists to measure whether AI can perform real life science research decisions — and even the strongest models clear only about a third of the tasks.
Google's June 2026 Pixel Drop delivers Android 17 alongside two generative AI features: Gemini Omni for video creation and Lyria 3 for AI-powered music composition.
Zhipu AI's Z.ai released GLM-5.2 on June 13 — a 744B-parameter MoE model with a functional 1-million-token context window, two thinking-effort levels, and MIT-licensed weights promised the following week.
MiniMax launched M3 on June 1 — an open-weight model using a new sparse attention architecture (MSA) to deliver a 1M-token context window, native multimodal input, and SWE-Bench Pro scores above GPT-5.5 at 5–10% of the cost.
Launched at Google I/O 2026, Gemini 3.5 Flash surpasses the previous-generation Pro on coding and agentic benchmarks while running 4x faster at 40% lower cost.
OpenAI replaced its default ChatGPT model with GPT-5.5 Instant, cutting hallucinated claims by 52.5% on high-stakes prompts. On June 9 the companion personalization feature — memory of past chats, files, and Gmail — expanded from paid tiers to all Free and Go users.
Apple unveiled a Gemini-powered Siri overhaul at WWDC 2026 alongside iOS 27 Extensions, a framework that lets users swap in Claude, ChatGPT, or Grok as their default AI across Apple Intelligence features.
Google unveiled Gemma 4, an Apache 2.0 open model built for advanced reasoning and agentic tasks, alongside TurboQuant — an ICLR 2026 algorithm that sharply cuts the memory overhead of KV caches in LLM inference.
Meta reversed its open-source strategy with Muse Spark, its first proprietary LLM built after a $14.3 billion, nine-month rebuild of its entire AI stack under Superintelligence Labs.
AIWire is launching a daily-updated section that distills the latest LLM news in both Korean and English.