AI Models & APIs
-
FLUX 3 x mimic — video-action model for robots: Audi is already testing it on the line
The company that made FLUX the king of image generation just pointed the same architecture at robot arms. What it actually is FLUX 3 x mimic is a video-action model. Black Forest Labs took the FLUX 3 multimodal backbone and fused it with robotics startup mimic’s manipulation learning. The result is one model that watches… Continue reading
-
Echo (TracerML) routes across GLM-5.2 and Kimi K2.7 to hit Fable-level quality at a third the cost
Echo isn’t a model. It’s an inference orchestration API from Adam Rida’s TracerML that sits in front of a pool of open-weight models — GLM-5.2, Kimi K2.7 and others — and decides, per request, which ones answer, how much compute each gets, and how to merge their outputs. Swap your single closed-model endpoint for Echo,… Continue reading
-
FLUX 3 (Black Forest Labs) beats Runway Gen-4.5 in 77% of video preference tests
Black Forest Labs, the team behind the FLUX image models, just shipped FLUX 3 — and it’s no longer just an image generator. It’s one model that produces image, video, and audio from a single set of weights, jointly trained instead of three separate models bolted together behind one API. What it actually does Text-to-video,… Continue reading
-
Upstage Solar Open 2 (Solar-Open2-250B): a 250B model that fires only 15B per token
Korea’s Upstage just put out Solar Open 2, an open-weight LLM rebuilt from scratch for agentic work — the multi-step, tool-calling, document-heavy stuff that eats context. It’s a Mixture-of-Experts: 250B total parameters, but only 15B activate per token. They reused just 2.3% of the last generation. This is a ground-up model wearing an old name.… Continue reading
-
Claude Opus 5: Polymarket bet 88% on a July 23 launch that hasn’t shipped
Claude Opus 5 is Anthropic’s next flagship LLM — a reasoning model you’d hit through the Anthropic API, not an app. Prediction markets got so confident about a July 23 debut that Polymarket odds briefly touched the high 80s. It’s July 24, and the model still isn’t live. Classic hype-cycle whiplash. What the leak actually… Continue reading
-
Kronos — open-source foundation model for financial markets hits 32.9k stars, gaining ~400 a day
Someone finally built GPT for candlesticks. Kronos, from developer shiyu-coder, treats K-line charts (OHLCV bars) as a language: a dedicated tokenizer quantizes continuous price and volume data into discrete tokens, then a decoder-only Transformer predicts what comes next. Same recipe as an LLM, pointed at markets instead of text. Pre-trained on 45+ global exchanges, MIT-licensed,… Continue reading
-
Cactus Hybrid (Gemma 4) teaches a phone-sized model to say “I don’t know”
Most on-device LLMs fail the same way: they’re confidently wrong and have no idea. Cactus Hybrid fixes the “no idea” half. It’s a post-trained version of Gemma 4 E2B with confidence probes baked directly into the checkpoint — every answer comes back with a score from 0 to 1 as structured data, so your app… Continue reading
-
GigaToken hits 24.53 GB/s on GPT-2 — 989× faster than HuggingFace tokenizers
Everybody assumed tokenization was already solved because the baseline was written in Rust and multithreaded. Marcel Røed spent a few days proving it was leaving three orders of magnitude on the table. HuggingFace tokenizers does 24.8 MB/s on GPT-2. GigaToken does 24.53 GB/s on an AMD EPYC 9565. Phi-4: 24.00 GB/s, 801×. Llama 3: 22.15… Continue reading
-
White House accuses Moonshot AI of distilling Anthropic’s Fable to build Kimi K3 — 3.4 million Claude exchanges, one week before the weights open
Michael Kratsios, director of the White House Office of Science and Technology Policy, named Moonshot AI on Wednesday. His claim: the administration has information that Kimi K3 was built with large-scale distillation of Anthropic’s Fable, run through a purpose-built internal platform that could switch between access methods to avoid detection. What Kimi K3 is An… Continue reading
-
Laguna S 2.1 (Poolside): 118B params, 8B active, trained from scratch in nine weeks
Pre-training started May 22 on 4,096 H200s. The model shipped July 21. That’s under nine weeks from zero to a coding model that beats Qwen 3.7 Max on SWE-Bench Multilingual. Laguna S 2.1 is an open-weight MoE model built for long-horizon agentic coding — 118B total parameters, 8B active per token, 1M context. Not a… Continue reading
Top AI Product 