NVIDIA ships Cosmos 3: open physical-AI foundation models in three sizes
NVIDIA released Cosmos 3, an open-source physical-AI foundation model family (64B Super, 16B Nano, 4B Edge) that unifies vision reasoning, world generation, and action prediction in one architecture rather than separate pipelines. The models rank first on several open-weights benchmarks — including Artificial Analysis's open-weights text-to-image board and RoboLab's robot-policy benchmark — and are built around the problem NVIDIA frames as central to physical AI: "every deployment is a specialization problem," so teams use Cosmos 3 to generate synthetic training data, simulate future environments, and adapt one base model to robots, autonomous vehicles, and vision systems. The release lands alongside an expanded Cosmos Coalition that now includes major Japanese manufacturers — a push for an open embodied-AI ecosystem rather than a single closed frontier model.
Sources & depth
AMD acquires Taalas, betting on chips that etch models directly into silicon
AMD is acquiring Taalas, a Toronto-based startup (founded 2023) that builds "model-specific integrated circuits" — chips with model weights etched directly into silicon rather than loaded from memory. Taalas's HC1 chip reportedly hit 16,960 tokens/second on Llama 3.1 8B, about 48x an equivalent NVIDIA GPU, but the approach trades away flexibility: switching to a different model means re-fabricating the chip, so it only makes sense for large infrastructure providers and model developers who can commit to a model long enough to justify the mask cost. AMD plans to fold the technology into its accelerator roadmap alongside Instinct GPUs and ROCm; neither company disclosed deal size or independently verified performance numbers.
Sources & depth
Anthropic cuts Fable 5's biology-safeguard false positives by ~85%
Anthropic refined the biology classifiers that gate Fable 5's dual-use-risk responses, cutting false positives by roughly 85% so the model can answer everyday health and educational questions — like interpreting a lab result — that it previously over-blocked. The company frames the underlying problem as genuinely hard to automate: vaccine-development knowledge and dangerous-pathogen knowledge look similar to a classifier, so the fix narrows what counts as "professional drug development or dangerous biological research" rather than loosening the guardrail itself. It's the latest step in Anthropic's stated approach of expanding access gradually through safeguard refinement and trusted pathways, not a one-time capability unlock.
Sources & depth
DeepMind's WeatherNext gains a full day of cyclone-forecast lead time
Google DeepMind says WeatherNext's three-day tropical-cyclone forecasts (track, intensity, wind structure) now match the accuracy prior models could only deliver two days out — DeepMind calls it roughly a decade's worth of forecasting progress in one model generation. The model adds more than 24 hours of lead-time advantage and materially lower 3-day position error (around 100km versus competing models), and gets there while running on coarser 28×28km input resolution instead of the high-resolution grids traditional numerical weather models need. More lead time on cyclone track and intensity converts directly into earlier evacuation and preparedness windows — the practical payoff behind the accuracy number.
Sources & depth
Unsloth ships DSpark-default GGUFs for DeepSeek V4-Flash, adds Kimi K3
Unsloth's v0.1.523/524-beta releases make DSpark speculative decoding the default path for its DeepSeek-V4-Flash-0731 Dynamic GGUFs — about 2x faster inference per the release notes — and add four quantization tiers, from an 83GB UD-IQ1_S up to a lossless 162GB UD-Q8_K_XL, so the model fits anything from a single 128GB machine to a full-precision multi-GPU rig. The same release adds day-one Kimi K3 support plus download-reliability fixes (automatic HTTP fallback when XET fails, better handling for Colab and low-memory machines). Nothing changed about the underlying V4-Flash weights — this is a tooling-maturity update, not a new model — but running V4-Flash locally just got meaningfully cheaper and faster to set up.
Sources & depth