Black Forest Labs releases FLUX 3 Image Klein, a 5.6B-parameter open weights model for text-to-image generation and editing with up to 10 reference images. The company says it tops its internal evaluations against larger open models. The FLUX Kommunity License permits free commercial use below $5M annual revenue, and API pricing starts at $0.009 per 1K image.
NVIDIA releases Nemotron 3.5 Super VL, an open 120-billion-parameter mixture-of-experts multimodal model that activates 12 billion parameters per token for agentic work. Scale AI says post-training it for Skydance credit processing cuts inference cost 40 times. Ramp claims 10x lower latency for receipt parsing.
Meta releases Muse Spark Open, an open-weight 816B-parameter multimodal model it says is the most capable from the U.S., tuned with Apple to run locally on two M5 Ultra Mac Studios with 1TB combined memory. It scores 39 on Artificial Analysis' Intelligence Index, behind China's GLM 5.3 at 45.
meta-models posted Muse Spark Open, an 816-billion-parameter model with 42 billion active parameters, to Hugging Face in BF16 and NVFP4 weights. The collection lists two image-text-to-text repositories, one at 822B and an NVFP4 version at 484B, both updated within the past day.
Facebook Research's AudioCraft library ships MusicGen, AudioGen, EnCodec, and the JASCO text-to-music model, which conditions on chords, melodies, and drums. The code runs on Python 3.9 and PyTorch 2.1.0 and carries an MIT license, while model weights fall under CC-BY-NC 4.0. The repository lists 23.7k stars.
Cheng Lou publishes Handwired, a tool that rewrites regular JavaScript into matrix multiplication plus ReLU so programs stay differentiable without training, and it runs on JS or WebGPU. The MIT-licensed repository has 1.4k stars, and a demo claims JSON search over 100,000 documents in milliseconds.
Perplexity open-sources SPACE, the sandbox platform behind its Computer agent, saying it replaced its commercial sandbox provider in July. The company says SPACE handles millions of sandbox creations daily and that testing found no virtual machine escape in 108 attempts. Akamai has launched SPACE on its cloud platform.
Open-weight models process 71.5% of AI Gateway tokens and 23% of spend in September, Vercel reports, with their spend share quadrupling since June. TypeSafe AI's Jev was the fastest-adopted model on record at 35% of paid teams in two weeks, yet took just 0.27% of spend.
A fork of PS5LM called PS5LM Plus runs llama.cpp on a jailbroken PlayStation 5's GPU, developer abhi245y says, loading GGUF models such as Qwen 3.8 27B at about 20 tokens per second inside an 11.44 GiB heap. It also serves an OpenAI-compatible API and chat UI over the network.
DigUp ships Windows and Linux beta builds of its on-device semantic file search, joining the existing Mac app. The open source tool runs Google DeepMind's EmbeddingGemma 2, a 740M-parameter model, to find photos, PDFs, audio and video by meaning. The 865 MB model downloads once and stays local.