A focused feed of models, agents, research and open-source releases for people building with AI.
Phonon is a free, local voice-typing app for Apple Silicon that targets Wispr Flow parity Refiner is a LoRA fine-tune of LFM3-1.7B , trained with DeepSeek-generated targets Author measures 65 ms key-up latency on M4 Max versus ~520 ms for Wispr Flow Claude Opus judge preferred Phonon's output 90 to
STAND, a minimalist supervised Bi-LSTM baseline, beats 20 anomaly detection methods across 5 benchmarks ( paper ). Just 5 to 7 labeled failure episodes are enough to surpass complex unsupervised deep learning models.
EfficientViT-L2 is an MIT Han Lab vision backbone using multi-scale linear attention instead of softmax. The r288_in1k checkpoint has 63.7M params, 11.5 GMACs, 288x288 input, Apache-2.0 license. Delivers up to 13.9x GPU latency reduction over SegFormer on Cityscapes with no accuracy loss.
Artificial Analysis launched Endpoint Accuracy Index v1.1 , finding nearly 1 in 4 serverless endpoints score below reference. Index composites BFCL v4-500 (tools), HLE-250 (reasoning) and AA-LCR-25 (long context) against a self-hosted baseline.
Antigravity Remote Agents run sessions in Google-managed cloud VMs, independent of your local machine Attach GitHub repos, Google Drive files, and uploads as Resources in a single project CLI access via agy --remote-agent flag, requires version 1.4.12 or later Default permission preset is turbo
Stanford's DSPy+HELM integration lifts benchmark accuracy 6% on average by swapping static prompts for structured ones. Leaderboard rankings flip on 5 of 7 benchmarks when models are evaluated near their prompt-tuned ceiling.
Thinking Machines released Inkling 2 Small , a 594B MoE with 17B active parameters and image input. Scores 31 on the Artificial Analysis Intelligence Index, with gains in coding, vision, and knowledge work.
FLUX 3 Image Klein ranks #2 open-weights on Artificial Analysis T2I at Elo 1043, one point behind Qwen-Image-2.1. 5.6B parameters, runs on 8GB VRAM consumer GPUs with INT8 quantization. Priced at $9 per 1,000 images on BFL API, 40% cheaper than FLUX.2 [klein] 9B.
Nous Research released Hermes Agent v0.23.0 , The Hearth Release, consolidating roughly 1,650 PRs since v0.22.0. The gateway can now serve the full Desktop workspace in a browser tab, no native installer needed.
Vietnamese sentence embedding model combining SimCSE contrastive learning with PhoBERT backbone, 135M parameters. Crossed 850K monthly downloads on Hugging Face , used by 22+ public Spaces. Scores 69.46 on ViRerank and 87.74 on MNLI-R, leading Vietnamese reranking benchmarks.
Cursor now builds and deploys live web apps from a prompt, with database, auth, and hosting handled automatically. Cloud agent scaffolds the app, spins up Postgres, configures Cursor SSO, and provisions hosting on request.
Krea launched a rebuilt Video workspace unifying generation, editing, style transfer and sound in-browser. Supports Seedance 2.5, Veo 3.1, Sora 2, Kling, Wan and Runway behind one prompt box. Ships cinematic camera controls, 100+ viral presets and UGC-ready ad templates.
Ampersand's multi-agent 'Co-Coding' mode hits 92.2% on SWE-bench Pro v2 HARD-51 , matching Fable 5.1 and Gemini 4 Argon. Scores 77.9% on DeepSWE v1.1, roughly 70% cheaper than Fable 5.1 at comparable quality.
World Labs released Frameboy , a browser demo streaming Atlas world model frames at up to 15 fps. Runs a speed-optimized variant of Atlas on cloud GPUs, preserving novel view synthesis quality. Scenes can be seeded from a single image or multiview captures of real locations.
METR estimates Anthropic's model capabilities are now improving 2x faster than the March 2024 to October 2025 baseline. AECI benchmark scores accelerated 1.9x since October 2025 (95% CI: 1.5-2.8x), robust across most methodological choices.
Meta launched Muse Realtime Avatar in the Muse app (US, 18+), animating reference images live during voice chats. Audio-driven Diffusion Transformer streams 448x768 video at 25fps with ~870ms end-to-end latency.
Microsoft Research released PDD Kit , a toolkit that formally verifies agent-written Rust code against developer intent. It converts plain-language intent into Verus specs and proves code meets them for every input.
New framework enables safe quadruped locomotion on uneven terrain using only IMU, encoders, and force sensors. Couples terrain mapping, contact, and state estimation into one probabilistic loop instead of three separate modules.
Anthropic launched Personal Finance in Claude , a beta hub for asking questions about linked bank and card accounts. Rolling out this week to US users on Pro, Max, and Ultra plans. Access is read-only via Plaid; Claude cannot move, manage, or invest money.
Runway integrates Gemini Omni 1.2 Flash for in-app generation and editing Clips extend to 15 seconds, up 50% from Omni 1.1 Flash Pricing drops 30% to $0.07 per second for 720p video Multi-turn conversational editing lets you refine without regenerating Supports 4K upscaling, reference images, and
Nanbeige LLM Lab released Nanbeige4.3-33B , a dense 33B reasoning model under Apache 2.0. Scores 20 on Artificial Analysis Intelligence Index , matching Qwen3.8-27B at similar size. Leads Terminal-Bench 4.0 at 23.8 versus 6.3 for Qwen3.8-27B, with strong agentic coding results.
Decart released Lucy 2.6 , its realtime video editing model, now the default behind lucy-latest. Upgrade targets character identity, clothing fit, and product fidelity from reference images during live streams. Runs at 720p, 30 fps realtime and 20 fps batch, at $0.02/sec (or $0.04/sec fast mode).
Goodfire introduces mechanistic shaping , adding interpretability-based penalties to the training loss. Two variants: reliance shaping penalizes using a concept, expression shaping penalizes representing it. Pushed RL agent honesty from 0% to 96-98% on a reward-hacking Countdown task.
Devin Review 2.0 ships in Devin Cloud and Desktop with autofix enabled. Per-PR review cost down 82% since August, typically $0.25-$0.75. Average cost per bug fixed is around $2.25 across Cognition's customers. Median latency improved from 139 to 94 seconds; bugs found per PR rose to 1.32.
No stories match these filters.
