Thursday, October 15, 2026

AidenNVIDIA today released Nemotron 3.5 Super VL, an open 120-billion-parameter model that reads images and documents, plus a small local model.

AidenJonah, our AI products analyst. What's actually out today?

JonahSuper VL ships with open weights, open datasets and training recipes. It's on Hugging Face, OpenRouter and ModelScope now, and as an NVIDIA NIM microservice. NVIDIA says it's strongest at OCR, tables, charts and documents like PDFs.

AidenNVIDIA also claims up to 5x the throughput of the next-best model in its class. And the small one?

JonahThat speed claim is NVIDIA's own. The small one is Bolt 74B-A9B, also on Hugging Face. Native 2-bit quantization fits it in 24GB of memory, for RTX PCs and workstations, and it works with llama.cpp, Ollama and LM Studio. A 4GB Bolt for voice assistants is promised for October 23.

AidenSo, for you: if you work with documents and charts, the Super VL weights are on Hugging Face today. If your GPU has 24GB, Bolt is built to run locally.

AidenDailyChat is AI-generated, with AI characters. Not professional advice. See you tomorrow.

AidenMeta says medical researchers using its Muse AI models drafted 10 health research manuscripts in about 4 weeks.

AidenLena, our AI for science analyst. How did that work?

LenaMedical researchers used an internal coding tool on Muse Spark 1.3 and Muse Wonder, with 5 million de-identified health records. The AI wrote and ran all the code. The experts set the questions and made the scientific decisions.

AidenParticipants reported up to a 10x productivity gain; some drafts took 7 days. That's Meta's own report, though.

LenaYes, and that's the best case. Each paper got feedback from the cohort, a biostatistician's review and summaries from independent experts. Meta doesn't say any has passed journal peer review.

AidenSo how solid are the findings?

LenaEarly. The studies looked back at existing records, so they can't show cause and effect. One reviewer called a paper hypothesis-generating; another said one overstated its mortality claims.

AidenSo, for you: if you work in health research, the post we link has all 10 manuscripts. Read them as early drafts.

AidenDailyChat is AI-generated. Aiden and Lena are AI characters, not real experts. Facts come from the sources we link. Not medical or professional advice. See you tomorrow.

AidenMistral today released Timestral, its first family of open models for forecasting data over time, in 4 sizes.

AidenJonah, our AI products analyst. What's actually out?

Jonah4 models, from 5 million to 209 million parameters. The weights are on Hugging Face under a modified MIT licence. They forecast series like sensor readings, and give a range of likely values across 21 quantiles.

AidenMistral says the largest one ranks first among pretrained zero-shot models on 3 benchmarks: GIFT-Eval, TIME and fev-bench. How much weight does that carry?

JonahThose are Mistral's own results, and a technical report with the full numbers is still to come. What you can check today is speed. It runs on a CPU, no GPU needed. Mistral says a forecast takes about 140 milliseconds with the largest model on an 8-thread server CPU, and 10 with the smallest.

AidenSo, for you: if you forecast from time-series data, like sensor streams, the weights are on Hugging Face today. Test them on your own data first; Mistral itself notes real data can differ from benchmarks.

AidenDailyChat is AI-generated, with AI characters. Not professional advice. See you tomorrow.

AidenMeta today released Muse Spark Open, an open-weight multimodal model with 816 billion parameters.

AidenJonah, our AI products analyst. What's actually out today?

JonahThe weights, under Meta's Muse Model License, which has limited commercial terms. They're on Hugging Face after an access form, and on SageMaker, Google Cloud and Microsoft Foundry. Hosted APIs are live at 7 providers, including Groq, Together AI and OpenRouter.

AidenMeta calls it the most capable US open-weight model. It has 42 billion active parameters and a 1-million-token context. Can I actually run it myself?

JonahThat most capable line is Meta's own, and its full evaluation report is still to come. Running it yourself is heavy. The model card puts it at a minimum of 700GB of GPU memory for the NVFP4 weights. Meta says it also runs on 2 Apple M5 Ultra Mac Studios with MLX.

AidenSo, for you: if you build with open models, the hosted APIs are the quickest way to try it today. Read the license terms before you ship anything commercial.

AidenDailyChat is AI-generated, with AI characters. Not legal or professional advice. See you tomorrow.

AidenBlack Forest Labs today released FLUX 3 Image Klein, an open-weight image model with 5.6 billion parameters.

AidenJonah, our AI products analyst. What's actually out today?

JonahBase and distilled weights are on Hugging Face; the base model is for fine-tuning. It makes images from text, edits them, and works from up to 10 reference images. The distilled one is also in the Black Forest Labs API.

AidenIt's under the FLUX Kommunity License: free commercial use under $5 million in yearly revenue. Black Forest Labs says it ranks first among open-weight image models. Can I run it myself?

JonahThat ranking is from Black Forest Labs' own evaluations. An INT8 version runs on GeForce RTX cards with as little as 8GB of memory. In tests Black Forest Labs publishes, the distilled model takes about 4 seconds per one-megapixel image on an RTX 5090, in ComfyUI.

AidenSo, for you: with an RTX card of 8GB or more, you can try the weights today. Companies above $5 million a year should read the license first.

AidenDailyChat is AI-generated, with AI characters. Not legal or professional advice. See you tomorrow.