ModelsNews
Adventists AI released the Sophon-Realtime v1 voice model family under Apache 2.0. The 32B model listens continuously, decides when to speak, picks up tone of voice and can add a laugh or a sigh to its reply. The Tag and Flash builds run on a single 80 GB GPU.
· 3 min read
ModelsNews
Thinking Machines Lab released the first model of its second generation. Inkling 2 Small is a 594-billion-parameter MoE with 17 billion active parameters per token and image input. It runs on Tinker with closed weights for now; Apache 2.0 weights are promised after external safety testing.
· 3 min read
ModelsNews
Ricursive Intelligence released Ripple-65B Preview under Apache 2.0: a 65-billion-parameter dense model with 1M-token context, pretrained on 12 trillion tokens. Its own model card reports 80.5 on Terminal-Bench 2.1, above US open-weight MoEs 8 to 15 times its size.
· 3 min read
ModelsNews
Nous Research released Arete-27B, a supervised fine-tune of Alibaba's Qwen3.8-27B meant to reinforce Western values on a Chinese-trained open-weight model. On Nous's 803-prompt eval, answers containing propaganda fell from 22% to 0.9%, and the fact score rose from 2.32 to 2.98, with no meaningful capability regression on the tests they ran.
· 4 min read
ModelsNews
Ant Group released Ling-3.2-flash, a native multimodal MoE with 560B total / 25B active parameters, a 1M-token window and text, image and video input. Chat plus OpenAI/Anthropic-compatible APIs are live, free for two weeks. Weights stay closed.
· 1 min read
ModelsNews
Moonshot AI's Kimi K3.1 Pro Highspeed is now on Amazon Bedrock for agentic coding and document work at high output tokens per second. It keeps a 1M-token context, explicit prompt caching and cross-Region inference.
· 1 min read
ModelsNews
Meta is rolling specialized versions of its models, trained with Muse Wonder, into Muse and Meta AI. The Personal Superintelligence Index rises from 38% to 47%, with Active Memory and Active Focus for fewer forgotten goals and stronger recovery.
· 2 min read
ModelsNews
NVIDIA released four new 16B-parameter Cosmos 3 Nano open-weight checkpoints for interactive world simulation: multi-sensor driving data, streaming observations, simulator-to-RGB transfer, and bimanual robot action-to-video.
· 2 min read
ModelsNews
OpenAI's GPT-6 Cyber scores 48.5% on ExploitGym, found 81 new Firefox vulnerabilities and almost always stays within its authorized scope.
· 2 min read
ModelsNews
Google released Gemini Omni 1.2 Flash, a video model with 15-second clips, multi-turn conversational editing and tighter reference adherence. The price drops 30% to $0.07 per second for 720p, and for the first time Omni is generally available on the Gemini API and Google Cloud.
· 2 min read
ModelsNews
OpenAI shipped Images 2.6 in ChatGPT, Codex and the API: faster generation, GIF animation, tighter editing, plus new developer controls.
· 2 min read
ModelsNews
OpenAI's cybersecurity model GPT-5.6 Cyber is now available to approved customers in the AWS GovCloud (US-East) region through Daybreak Red on Amazon Bedrock.
· 2 min read
ModelsNews
Dharma-AI's 2-billion-parameter model turns a document image into structured JSON, optionally following a schema you define. On the company's own benchmark it scored 0.834, ahead of every frontier model it tested.
· 3 min read
ModelsNews
India's Sarvam released Prism, which returns calibrated probabilities for every option in a single step. It is built for routing and checks in agents.
· 2 min read
ModelsNews
Pruna turned the MiniMax H3 video model into an image model that is 7x smaller and 5x faster. Its license, however, excludes the EU.
· 2 min read
ModelsNews
Tencent released the final Hy4: its first Hy model with vision and computer use. API pricing stays the same, and open weights are promised for later.
· 2 min read
ModelsNews
China's Deep Principle released Hydrogen-1.0, an MIT-licensed model tuned for chemistry, materials science and life sciences. Its own tests show it beating GPT-6-Astra on a research benchmark.
· 2 min read
ModelsNews
SenseTime has released SenseNova-Loop-9B-Preview under Apache 2.0. The 9B multimodal model runs its middle 24 layers twice with shared weights, and the company says this delivers 27B-class results on agentic tasks.
· 2 min read
ModelsNews
Nace.AI has released Drex Max v1, its most capable decision model, as open weights. The 27-billion-parameter model does not generate text. It returns a probability for every option in one step, and its quantized version runs on a 24 GB Mac.
· 2 min read
ModelsNews
Samaya post-trained a 30B Nemotron model with 3B active parameters for financial research. It scores 58.9% on Samaya's FrontierFinance benchmark, above Claude Fable 5.1's 55.9%, at roughly an eleventh of the cost.
· 2 min read
