dsh-Go Brings DeepSeek's Coding Agent to Android Without Root Access
A Flutter-built Android client packs the DeepSeek Harness coding agent into a phone, with a built-in Linux sandbox and remote control for your desktop.
20 stories
A Flutter-built Android client packs the DeepSeek Harness coding agent into a phone, with a built-in Linux sandbox and remote control for your desktop.
A JAX-native multi-objective RL framework trains humanoid robots across six conflicting goals in minutes, delivering 21 to 270 times speedups.
Runway's creative Agent now writes and executes code, opens files, and browses the web to deliver finished brand kits, ad variants, and motion graphics.
Zyphra's new hierarchical attention trades the quadratic cost of self-attention for a tree-structured search, projecting hundreds of times speedup at billion-token contexts.
Pydantic V3 is in development with a gentler migration story, V1 reaches end of life, and a paid enterprise support tier arrives for production teams.
Factory folds Typedef's data context layer into its Droid coding agents, aiming to give enterprise AI engineers lineage-aware migrations and pipeline debugging.
Google's Gemma team showed an untethered Reachy Mini robot that chats, reasons, and drives itself using only on-device models over LiteRT.
Lovable's Perplexity connector now works without an API key, letting any app pull cited, real-time web answers via managed billing.
A community Windows desktop app packages Boson AI's Higgs Audio v3 TTS into GGUF quantizations with a local C++/CUDA engine for offline voice cloning.
AutoTrust's open 9B model packs calibrated System 1 decisions and full System 2 reasoning into one Qwen3.5 backbone, matching a closed teacher at a third the latency.
Krea's node-based workflow editor now supports real-time multi-user editing, letting teams build and run AI pipelines together on one canvas.
humans& trained a tutor model against a simulated learner called Persimmon, topping a real-classroom benchmark and beating Claude Sonnet 5.5 and GPT-6 Astra.
A new paper splits task vectors into Task Recognition and Task Learning parts, then tunes their weights to beat the originals.
Maya Research dropped open weights for a streaming TTS model covering 11 Indian languages plus English, with 20 voices and zero-shot cloning under Apache-2.0.
Microsoft's 4B hybrid agent hits 66% on OSWorld-Verified by learning when to click a button and when to open a shell.
DART tracks a drifting ID vs OOD axis at test time, cutting false positive rate by 49 points on corrupted ImageNet benchmarks.
A LoRA-only recipe converts frozen Gemma, Llama, Qwen, and Ministral checkpoints into masked diffusion language models in under four GPU hours.
OVHcloud's AI arm released Olala mini, an Apache 2.0 hybrid Mamba-3 and attention MoE that fits on a laptop and reasons before answering.
Cagliostro Lab's open-source anime SDXL finetune ships with 8.4M training images, a tag-based prompt grammar, and a refined Opt variant for better anatomy and color.
Openjutsu brings Higgsfield-style one-click character and object swaps for video to local GPUs, running MiniMax H3 and SAM 3.1 through ComfyUI.