Model & Agent Releases
Spark-X2.6-4B on-device. A compact 4B model for on-device use with improved coding, agent, long-context, and reasoning through post-training.
Ornith-1.5 self-improvement. Ornith-1.5-122B extends end-to-end self-improvement to jointly optimize task generation, scaffolds, and rollouts, built on Qwen3.5 and Gemma4.
Agnes-3.0-Qwen agentic MoE. A ~180B multimodal MoE with 1M context shows strong agentic results on τ³-Banking and Terminal-Bench, and outperforms GPT-5.6 Sol and Claude Opus 4.8 on IFBench.
Real-time voice stacks. Mistral’s Voxtral 3 and ElevenLabs’ Eleven Realtime both ship speech-to-speech models for real-time agents, with transcription, TTS, and SDKs; Eleven Realtime Max scores 98.8% on Big Bench Audio with 0.57s response.
Small open TTS. Phonon open-sources 31M/40M TTS models for edge devices, while Pocket TTS shrinks voice cloning to 1.9M parameters.
Tools, Frameworks & Inference
TorchTPU open source. Google open-sourced TorchTPU, a native PyTorch backend for Cloud TPUs co-developed with Meta, replacing PyTorch/XLA.
DeepInfer inference engine. DeepSeek open-sourced DeepInfer, the serving stack behind V4.1-Flash, plus deepseek-recipe for compatible self-hosted token processing.
Apple Silicon speculative decoding. New methods KFlash and Eagle-4 speed up local inference on Apple Silicon; KFlash benchmarks on MacBook Pro M4 Pro show ~65 tok/s for Qwen 3.8 27B.
- KFlash: Benchmarks of newly released versions of Apple Silicon optimized engines: 1.4.0 Splash (DFlash2 ~ 50 tok/s) vs 0.2 Kapitsa (KFlash ~65 tok/s), MLX(MTP) & more on Qwen 3.8 27b/4b/35b 4Q on Macbook Pro M4 Pro 48 GB 2024, 48GB & Macbook Pro M1 Pro 16GB 2021(opens in a new tab)
- Eagle-4 Released: Faster than DFlash2 & DSpark(opens in a new tab)
RK3588 NPU runtime. An open-source runtime without proprietary libraries runs Qwen3.5-4B about 2x faster than vendor SDK on Rockchip NPU.
Legacy GPU inference. Pantheon inference engine in plain C runs 35B MoE models on a Tesla K80 at ~21 tok/s, enabling cheap old hardware for local LLMs.
MCP criticisms. Figma’s Dylan Field questions MCP’s asymmetric context sharing while the company adds 20 new MCP partners.
Research Highlights
FairStress benchmark. FairStress measures demographic fairness in LLM decisions beyond final choice, examining how readily a model reaches decisions across demographic swaps.
ALTK-Evolve memory. Learned memory from training-task attempts pushes a GPT-4.1 ReAct agent to 67.5% on τ²-bench retail, topping the leaderboard without frontier models.
Scaling memory with n-grams. COUNT builds n-gram embeddings from host-memory counts to scale memory use, improving validation loss as much as 57% more compute at 1B scale.
REMORY residual memory. REMORY supplements summaries with soft memory tokens, improving source attribution and agent performance with only 5.2% input positions.
Surgical world model. NVIDIA Cosmos3-H-Dreams generates real-time surgical simulation from clinical recordings, running interactively on a single RTX PRO 6000.
Industry & Business
SEA digital economy. Southeast Asia’s digital economy to hit $368B GMV in 2026, with generative AI used by 243M people and agentic commerce driving a 60-fold surge in AI usage.
Anthropic cloud crunch. Anthropic struggles to secure cloud capacity from AWS/Google/Microsoft, which see it as competitor; leaked S-1 shows potential $750B cloud spend and $80.7B net loss in H1.
SpaceX AI profit. SpaceX’s AI business, now 52% of income, generated $7.5B revenue and $431M profit in Q3, led by Cursor, Grok, and neocloud.
OpenAI biotech acquisition. OpenAI acquires Orchestra Bio to strengthen Rosalind life sciences model and workbench; adds team and tech for drug-discovery agents.
Meta Muse downloads. Meta’s AI companion Muse reached 10.9M US downloads in 40 days, one of only 12 apps to hit that milestone.
Spotify AI slop. Spotify removed 50 million AI-generated slop/spam tracks from recommendations and changed system to favor human artists.
ByteDance chip orders. ByteDance, Alibaba, Tencent placed over $35B in Nvidia chip orders, including 3M RTX 5500 Pro, likely for inference and video gen.
Applications & Product Updates
Gemini in cars. Personal Intelligence comes to Android Auto and Google built-in, letting Gemini use Gmail, Calendar, Maps, Photos for driving assistance.
ChatGPT locked chats. Users can lock ChatGPT conversations with PIN/biometrics for privacy from shoulder-surfing.
Muse sharing paused. Meta paused Muse’s share feature after researchers found 6,800 publicly accessible links containing sensitive info like revenue and calendars.
Adobe eye opening. Lightroom mobile can open closed eyes in photos using generative AI with a single tap, replacing eyes realistically.
Local archive agents. A user built local AI agents to process a massive personal archive, generating connections and curiosities without frontier APIs.
Medical local models. A Lithuanian conference highlights growing use of local models in medicine due to strict patient data privacy.
Policy & Safety
Anthropic safety bill. Anthropic pitches AI safety legislation that could block model release if foreseeable risk kills 50+ people or causes $1B property damage.
US bans DeepSeek. US orders federal agencies to remove DeepSeek products and third-party services, citing national security and data concerns.
Meta TOS edit. Meta drops the word 'train' from commercial terms after backlash over AI training data scraping; says crawling terms only applied to ads/business products.
Altman’s endgames. Sam Altman reportedly pitched three AI safety governance options internally, including IAEA-like watchdog, lab consortium, or single-firm dominance.
Meta torrents legal. Court orders Meta to name executives who authorized torrenting pirated books, potentially exposing leadership in AI training copyright cases.
US/China open controls. Both US and China propose export controls on open AI models; China’s MOFCOM would require pre-approval for sharing frontier models internationally.
That's everything for this day – about a 5-minute read.