Open-Source AI

Today's Stories

1. 1

   NVIDIA's DIN Deploy ships C++ local AI samples

   NVIDIA published DIN Deploy, open-source C++ samples that combine ONNX Runtime with the TensorRT RTX execution provider, with support for Windows and Linux. Developers get a shared code path with vendor-specific CUDA code only in optional accelerated paths, so the same samples can run across systems that support the required tensor APIs.

   The samples are only as portable as the execution providers that support ONNX Runtime's tensor APIs. Watch the DGX Spark results, where whisper-large-v3-turbo ran 58.5x real time on GPU versus 3.8x on CPU.

2. 2

   DEER + GTF cuts RNN training time >100x for chaotic systems

   A NeurIPS 2026 spotlight paper reports that combining DEER with generalized teacher forcing (GTF) speeds up training of nonlinear RNNs on chaotic dynamical-system time series by more than 2 orders of magnitude (>100x). That is a claimed speedup of over 100x on the specific task of reconstructing chaotic dynamical systems from time series, suggesting training that previously took far longer may now be feasible, though this remains a paper report.

   Whether the more-than-100x speedup holds on other chaotic systems and tasks beyond those tested, and whether others replicate the DEER plus GTF combination. Watch the openreview links for updates.

3. 3

   AWS debuts open-source Strands Decider 2B decision model

   AWS released Strands Decider 2B, an open-source decision model that makes choices instead of generating text. Today's release is v.20, built on Qwen3.5-2B with a 1 million-parameter pointer head. By skipping text generation, the model avoids the token consumption and latency that slow down conventional LLMs, potentially speeding up agentic AI development.

   The model's usefulness hinges on whether developers adopt it for tasks like model routing and tool selection. AWS claims it achieves high accuracy on JevBench compared with other open-source 2B models.

4. 4

   Runway unveils Praxis-1, an open-weight robot AI model

   Runway announced Praxis-1, its first open-weight world action model for controlling robots, built on the same large-scale video pretraining as its world models. Runway says video pretraining gives a robot policy a head start over one trained only on robot action data, which could ease robotics' data bottleneck.

   Runway reported 16.1 cm placement error for web-video pretraining versus 16.0 cm for teleoperated robot video, a difference it calls not statistically significant.

5. 5

   Ideogram 4.5 edits only selected image areas, from 0.8 cents

   Ideogram launched Ideogram 4.5, which it says only touches the area users tell it to edit. It costs 0.8 to 22 cents per image across four quality tiers at native 2K resolution, and is live on the Ideogram platform and API. If the claim holds, creators editing a product shot or portrait can change one element without redoing the whole image. Ideogram says an open-weight release is coming soon, which could let others run a similar model themselves.

   The test is whether consistency holds after many edits, since even leading models still tend to produce artifacts after repeated changes. Ideogram has not said when the open-weight release will arrive.

6. 6

   Magnitude inference engine beats llama.cpp by up to 2x

   The Magnitude development team published an open-source inference engine that compiles and tunes model kernels on the PC itself, hitting about 2x llama.cpp's speed in certain conditions. Running open models locally for AI agents has leaned on engines built for datacenter batching or broad hardware support, so Magnitude's on-device tuning looks aimed at a gap those tools leave.

   The headline 2x figure holds only in certain conditions, so the test is whether the gains repeat as the team adds Expert streaming and broader multi-device optimization.