Thinking Machines' Inkling 2 Mini Loops Layers to Beat Larger Open Models
Thinking Machines releases a 26B looped-transformer MoE that runs on a laptop, posts 20.1 on AA Index, and ships with open weights for monitorability research.
20 stories
Thinking Machines releases a 26B looped-transformer MoE that runs on a laptop, posts 20.1 on AA Index, and ships with open weights for monitorability research.
OpenAI details its layered defense stack against adversarial distillation, claiming a drop in reasoning extraction success from 54% to 0.2%.
OpenAI's new medicinal chemistry benchmark grades frontier models on reading molecular structures, predicting potency, and picking the next compound to test.
Arena's refreshed Alignment Index shows OpenAI sweeping the top seven spots, Kimi K3.1 Pro leaping 28 places, and Anthropic's Claude Fable 5.2 debuting at #8.
Google's Gemini 4 Neon brings frontier reasoning to agent workflows at roughly half the cost of larger rivals, headlining a broader shipping spree.
Google's chat assistant can now generate short looping GIFs on demand using the Omni 1.2 Flash video model, replacing manual searches on sites like Giphy.
TurinTech's Artemis conversion shrinks Google's Gemma 4 E4B from 14.9 GB down to 4.22 GB while matching the full-precision model's next token 98.4% of the time.
Alibaba's Qwen3-14B gets an official FP8 release with fine-grained block quantization, cutting memory use while keeping dual thinking modes intact.
Kyutai's AURA puts a lip-synced face on live voice AI, streaming 576p talking-head video at 20.7 fps on a single H100.
Researchers show the Shapley value can both quantify and attribute group unfairness, cutting inference cost versus bootstrap on Census Income data.
Reson8's 0.5B open-weights speech-to-text model lands on Hugging Face under Apache 2.0, beating Whisper Large V3 at a third of its size across 45 languages.
A free, open textbook pairs a Python sandbox with foundational theory so newcomers can build and test their own negotiating agents from scratch.
Tencent's new flagship mixture-of-experts model packs 770B parameters, native vision, and a 1M token context window under an open license.
Kaggle and Eigent AI launch a benchmark where coding agents debug real RL training stacks, and the top model clears just 22.8 percent of tasks.
Sakana AI spins up an independent cyber defense lab in Tokyo, betting that quality engineering upstream beats patch-chasing downstream.
A new pair of unadjusted Langevin samplers cuts gradient evaluations by up to 17x on constrained distributions while staying stable at larger steps.
OpenAI patched an autonomous weapons alignment gap flagged by Vals AI, and the fix shipped in its newest frontier model Astra.
OpenAI's Responses API now preserves reasoning across turns when the conversation prefix stays intact, but edits silently drop prior thinking.
Cognition's FrontierCode 2.0 benchmark puts Google's new mid-tier model within a point of Claude Opus 5.5 while costing roughly a third per task.
A GGUF-quantized build of Google's UMT5-XXL encoder is quietly powering most local Wan video pipelines, shrinking a 22GB model to under 4GB.