Cartesia has launched Sonic-3.6, the latest model of its real-time text-to-speech mannequin. It arrives roughly three months after Sonic-3.5. The new change is naturalness, and this one is independently checkable. Sonic 3.6 now holds #1 on each Artificial Analysis speech leaderboards — 1,283 Elo on the Provider Voice board and 1,123 on the Controlled Voice…
Nous Research has shipped Bot Mode for Hermes Agent, its MIT-licensed open supply agent. Bot Mode replaces the single-agent session listing with a roster of named bots. Each bot is a actual Hermes profile, with its personal chat, reminiscence, abilities, and pinned mannequin. Bots message one another by means of a persistent Agent Inbox and…
ByteDance Seed and Tsinghua AIR have launched CUDA Agent, an agentic reinforcement studying system that trains a big language mannequin to put in writing GPU kernels that beat a compiler. The hole it targets is slim however cussed: frontier fashions already produce appropriate CUDA, they simply produce sluggish CUDA. On KernelBench, the bottom mannequin Seed1.6…
MiniMax launched MiniMax-Music3, an open-weights text-to-music mannequin. The mannequin takes two separate inputs: lyrics carrying part tags, and a detailed music description. It returns a full tune of as much as 5 minutes in a single technology, as 32 kHz, 16-bit stereo WAV. The structure pairs a Hybrid-LM, an 8B Global LLM with a 0.6B…
DeepSeek launched DeepSeek Harness v0.1 in developer preview and printed the complete supply code beneath the MIT license. The challenge ships as dsh at deepseek-ai/deepseek-harness. A harness is the layer between a mannequin and the atmosphere it acts in — the instruments, recordsdata, sandboxes, and management loop that allow an agent hold working. DeepSeek frames…
Z.ai simply launched GLM-5.3. GLM-5.3 runs on the identical 743B base mannequin as GLM-5.2. Every reported achieve comes from scaled post-training: extra job environments, extra atmosphere sorts, longer coaching. The outcomes land in two locations. Coding jumps most on the longest-horizon benchmarks, with Terminal-Bench 3.0 shifting from 4.6 to twenty-eight.3. Cybersecurity moved additional than Z.ai…
Cactus Compute has launched Needle 2, an open 45M-parameter mannequin for instrument calling, gadget use, and structured extraction. The complete mannequin ships as a single 14MB binary that runs a full session in about 28MB of RAM. Weights are educated and deployed at CQ2-bit utilizing Cactus Quants, and the mannequin is sealed inside the corporate’s…
Google has launched Gemini 3.7 Flash, the latest mannequin in its Flash tier, three weeks after Gemini 3.6 Flash. The model card describes it as a refinement of three.6 Flash with algorithmic enhancements to the core reasoning basis — not a brand new pretraining run. It accepts textual content, photographs, audio, and video throughout a…
Ask any AI engineer what ships quickest to a demo and slowest to manufacturing, and the trustworthy reply is nearly at all times the interface, forward of the mannequin, the retrieval pipeline, or the eval harness. A chat window that seemed intelligent within the pitch deck turns right into a bottleneck the second an actual…
Dyna Robotics has launched Dyna-2, a world-action mannequin for robotic manipulation. It was pre-trained on a couple of million hours of selfish human video. That is roughly 170 years of steady waking expertise. Robot studying has been bottlenecked by action-labelled information, which teleoperation should intentionally produce. Dyna-2 checks whether or not odd human video can…