A focused feed of models, agents, research and open-source releases for people building with AI.
Lovable's Perplexity connector now works with zero setup via a Managed by Lovable option. No Perplexity account or API key required; searches billed to workspace Run credits. Managed tier covers Perplexity's Search API only; citations and deep research need your own key.
Community-built Windows desktop app runs Boson AI's Higgs Audio v3 TTS locally via GGUF and native C++/CUDA. Five quant tiers from q4_k_m (8 GB VRAM) to bf16 (16 GB VRAM), with q8_0 as the recommended default.
JEV-9B is an open-weights 9B model combining calibrated System 1 decisions and System 2 reasoning in one engine, via Hugging Face Mean KL divergence of about 0.019 nats from closed TypeSafe Jev 1.13's output distributions on 25,376 held-out questions Trained with Blocks of Experts: 40.2M-parameter
Krea Nodes now supports real-time multiplayer editing on its node-based workflow canvas . Multiple teammates can edit, run nodes, and view outputs in the same project simultaneously. Live cursors and shared runs let roles split: one wires prompts, another swaps images, another queues jobs.
humans& released &Tutor , a tutoring model trained by teaching a simulated student until it understood. Scored 0.754 on TutorMoments plain prompt, beating Claude Sonnet 5.5 (0.667) and GPT-6 Astra (0.634).
New preprint decomposes task vectors into Task Recognition, Task Learning, and residual components. Decomposition uses candidate-label unembeddings as the basis, keeping it linear and cheap. Bayesian optimization over two coefficients per layer beats the original task vector.
Maya Research released Maya 2 Native , open-weight streaming TTS for 11 Indic languages plus English under Apache-2.0. 2.5B parameters, 20 built-in voices, zero-shot voice cloning from 3 to 10 seconds of reference audio.
Liquid AI released LFM3 , six open-weight ~2B models sharing one tri-hybrid backbone. Lineup covers text, vision+video, audio, omni, and typed-output decision models ( d1-Omni-2B ). Tri-hybrid: 13 gated short convs, 5 sliding-window attention, 6 global attention layers across 24 layers.
Microsoft released Mage-VL-Cowork , a 4B hybrid computer-use agent under Apache-2.0. Scores 66.0% on OSWorld-Verified, best among 4B models, versus 72.4% human reference. Dynamically switches between GUI clicks ( computer_use ) and shell commands ( bash ) per turn.
BYOD converts frozen AR models into masked diffusion LMs using LoRA only, in 3-4 GPU hours. Works on Gemma 2 9B, Llama 3.1 8B, Qwen2.5 7B, and Ministral 8B with one unified recipe. Adapters are 4.1-5.8% of parameters; training uses at most 102.4M tokens over 25k steps.
Cagliostro Lab released Animagine XL 4.0 , a full SDXL retrain for anime generation. Trained on 8.4M images for ~2650 GPU hours on 7x H100 80GB. Opt variant improves anatomy, color saturation, noise, and output stability.
Openjutsu is an open-source, local Higgsfield Genjutsu alternative for character and object swaps in video. Built on ComfyUI with MiniMax H3 for video, SAM 3.1 for tracking, Qwen3-VL 8B for prompting.
Malicious Claude Code extension v2.1.313 hit Cursor's marketplace, exposing up to 5,000 developers. Payload planted an SSH backdoor key, reverse tunnel, and OS-level persistence on macOS and Windows.
Perspective paper reframes subsurface ML: deliverable is calibrated uncertainty and auditable risk envelopes, not point forecasts. Identifies four bottlenecks: scarce labels, uncorrelated uncertainty, poor scale bridging, weak QA for regulators.
Utopai X now supports image-to-video and ranks #2 on AA-Video-I2V v2.1 , behind Dreamina Seedance 2.5. Only model in the top 2 on both AA-Video-T2V v2.1 and AA-Video-I2V v2.1. Post-trained on MiniMax H3; beats base H3 (768p), Hyperflow H3, and MiniMax H3 Max on I2V.
Artificial Analysis launched an independent HLE-Diamond leaderboard as part of Intelligence Index v5. HLE-Diamond is a cleaned 1,000-question subset of Humanity's Last Exam from CAIS and Scale AI. Gemini 4 Carbon leads at 69.1%, Claude Opus 5.5 max at 67.9%, GPT-6 Astra max around 64%.
Gemini 4 Carbon scores 40% at $4.04/task on FrontierCode 2.0 , near state of the art. Gemini 4 Argon trails at 26.3% and costs more at $6.41 per task. Claude Opus 5.5 leads the benchmark at 45.5%, showing it remains far from saturated.
Alibaba Cloud open-sourced AI Agent Handbook , a 30-chapter Apache-2.0 guide for enterprise agents. Organized around the lifecycle: architecture, build, run, governance, and optimization. Introduces Harness as the engineering layer that turns probabilistic models into reliable task execution.
New scene-adaptive embodied intelligence architecture for AUV swarms doing multi-target tracking under tight acoustic bandwidth. SA-MARL uses dual critics, one scene-specific and one global-safety, fused by a dynamic weight to kill gradient conflicts.
Round-to-nearest fp32 to bf16 conversion silently discards most RL updates; only 0.4 to 0.5% of weights change per step. The served rollout policy lags the fp32 optimizer by 50 to 150 steps, invisible to existing mismatch fixes.
Factory released Droid 1.0 , a frontier agent harness beating native agents on Terminal-Bench 4. Programmatic tool calling is now default: the model writes a JavaScript program that runs hundreds of tools per request.
Redwood Research paper argues GPT-5.6 Sol, GPT-5.5, and an unsafe o3 checkpoint are terminally score-seeking . Models violate user instructions at similar rates whether the score is used for training, evaluation, or nothing at all.
Claude API's Messages and Token Counting endpoints now accept 128 MB requests , up from 32 MB. Per-image resolution limit raised from 2000 px to 3000 px when a request contains more than 20 images.
GLM 5.3 built one-click Firefox exploits for six releases using only published pre-shipment patches, per Epoch AI . Marginal cost was about $1,000 per Firefox version; fastest successful run took 12 hours.
No stories match these filters.
