
Sakana AI Replaced Human Artists With AI and Exposed a Creativity Gap
Sakana AI replaced human users in the classic Picbreeder image-evolution experiment with VLM agents, revealing a fundamental gap in AI open-ended creativity.
20 stories

Sakana AI replaced human users in the classic Picbreeder image-evolution experiment with VLM agents, revealing a fundamental gap in AI open-ended creativity.

Unsloth's NVFP4 quants run Qwen3.6-27B on a 24GB GPU at 2.5x the speed, with MTP speculative decoding baked in

Kyutai and Mirelo release MuScriptor, an open-weight model that converts any multi-instrument recording into per-instrument MIDI using a novel three-stage training pipeline.

ElatoAI is an open-source platform that turns a $5 ESP32-S3 chip into a real-time voice AI device supporting 100+ models from OpenAI, Gemini, and more

Unsloth's NVFP4-Fast quant of Qwen3.6-35B-A3B delivers 15,636 output tokens/sec on a single GPU, beating NVIDIA's own release by 1.79x
A tiny 0.1B-parameter multilingual speech recognition model runs entirely on-device with a 200MB memory footprint on iPhones.

Alibaba's Wan-Streamer v0.2 runs full-duplex video conversations at 640×368 and 25 FPS with ~550ms total latency, all inside a single Transformer

Moonshot AI releases MoonViT-SO-400M as a standalone encoder, letting developers plug native-resolution vision into any multimodal pipeline

Moonshot AI open-sources Kimi-Audio-7B-Instruct, a single model that handles speech recognition, emotion detection, audio Q&A, and live voice chat — beating GPT-4o-mini on conversation quality.

Moonshot AI's Moonlight proves the Muon optimizer scales to 16B-parameter MoE models, cutting training compute in half while beating Llama and Qwen at the same scale.

Moonshot AI's Kimi-VL-A3B-Thinking packs 16B total parameters into a 2.8B-active MoE model that rivals 70B-scale VLMs on math and visual reasoning benchmarks
Moonshot AI's 1-trillion-parameter open-weight model tops the LMSYS leaderboard for open-source models, rivaling Claude and GPT on coding and agentic tasks

K2.7 Code HighSpeed exits beta and opens to all Allegretto+ subscribers, hitting 180–260 tok/s at 3x the credit cost

Datacurve's DeepSWE leaderboard now shows GPT-5.6 Sol, Terra, and Luna results — Sol tops all 13 models at 73%, with striking cost-performance gaps across the family.

NVIDIA's Flex-Forcing lets a single video diffusion model switch between fast streaming and high-quality generation at inference time, earning an ICML 2026 spotlight.

Google's LiteRT.js brings native on-device AI inference to the browser via WebGPU and WebAssembly, outperforming existing web runtimes by up to 3x

Cohere's open-source 2B Arabic ASR model tops the Open Universal Arabic ASR Leaderboard, beating Whisper v3 Large by 11 WER points with native dialect and code-switching support.

Robbyant open-sources LingBot-World 2.0, a 14B interactive world model that generates coherent, playable environments for over an hour without visual degradation

Google's SensorFM is trained on 1 trillion minutes of Fitbit and Pixel Watch data from 5 million people, transferring across 35 health prediction tasks with no task-specific architecture

Lance v8.0.0 lands with FM-Index substring search, vectorized RaBitQ, branch-aware versioning, and a sweeping index architecture overhaul