Google Magenta RealTime 2: open-weights real-time music at ~200ms latency, text/audio/MIDI control, 2.4B params — runs on a MacBook without a GPU. Previous latency: ~3 seconds.
MiniMax M3 offers 1M-token context, beats GPT-5.5 on SWE-bench-Pro, and releases open weights in 1-2 weeks — API now at $0.20/M tokens, a fraction of frontier pricing.
Ideogram 4.0 open weights are live — claimed #1 open image model, downloadable, fine-tunable on own data, available now at ideogram-ai/ideogram-4-nf4 on HuggingFace.
Google's Gemma 4 12B is encoder-free multimodal — text, audio, video, image — in 16GB VRAM under Apache 2.0. Day-0 in Transformers, llama.cpp, MLX, and Red Hat OpenShift.
NVIDIA's Nemotron 3 Ultra brings 550B open-source AI at 5x faster inference and 30% lower cost than competing open frontier models, now live on Hugging Face.
DeepSeek raising $3–4B at $50B valuation via China's national AI fund shifts the open-weights frontier to state-backed capital.
Qwopus3.6-35B: Qwen3.6 35B distilled on Claude Opus reasoning traces (71.9GB) brings Opus-quality reasoning patterns to an open-weight architecture.
Researchers release talkie-lm, a 13B model trained only on pre-1931 text: it defends luminiferous aether, can't arrange a sushi delivery, and raises the question of whether it can learn to code.
DeepSeek-V4's MIT-licensed 1M-context MoE and Kimi-K2.6's multimodal orchestration create the first complete open-weights agentic deployment stack.
DeepSeek V4 drops two open-weight models with 1M-context by default, CSA+HCA hybrid attention, and V4-Pro priced at roughly 1/7 Opus 4.7's output cost.
DeepSeek V4-Pro launches with 1.6T parameters, 1M context, and 10× KV cache reduction over V3.2 — multiplying inference concurrency roughly 10× on the same hardware.
Curated AI insights — sent when there's something worth your inbox.