Open source

AIGoogle says the Antigravity SDK now supports running agents entirely on a local machine with Gemma 4 and LiteRT. The post adds support for OpenAI-compatible endpoints, naming Ollama, llama.cpp, and vLLM as options for serving Gemma, and gives the install command pip install google-antigravity litert-lm.

Why it matters: The post names the specific runtimes and serving endpoints supported, letting developers judge whether their current local setup fits the new SDK path.

AIGoogle Antigravity says developers can now run open models such as Gemma 4 completely offline in its SDK. The setup uses Google AI Edge's LiteRT to run the model directly on a local GPU, with no API costs and no internet connection required.

AIInferact says its first TPU megakernel for Kimi K3 reaches 709 tokens/s on low-concurrency decode with DSpark speculative decoding, versus 450 tokens/s for its GB200 baseline. The company claims it is the first TPU inference megakernel, running the whole model in a single Pallas kernel, and says it is roughly 1.4 to 2x the GB200 baseline at batch sizes 1 through 8 without speculative decoding. Inferact says it is open-sourcing the kernel today.

AIBionic now includes a built-in interactive canvas where users can create Excalidraw diagrams that both they and Bionic can view and edit. The canvas supports collaboration on mockups, system designs, and process maps, and users can ask Bionic to implement what is drawn.

AIBlack Forest Labs says FLUX 3 Action is an open-weights 7B world action model that ranks first on the RoboLab benchmark. The company says it outperforms the previous best open model by 6.1 percentage points while using 56% fewer parameters and running up to 3.95x faster. The model predicts video and actions together, and the company is releasing the weights, code, fine-tuning recipe, benchmarks, and examples. It also integrated the model into Hugging Face's LeRobot with NVIDIA, with edge deployment on NVIDIA Jetson.

Why it matters: The release pairs benchmark results with the trade-off it claims to remove between world action model performance and VLA speed, which is useful context for robotics teams weighing open models.

AIDeveloper Stefano (@spedemo) built a 3D-printed, remote-controlled robot named Marvin that uses voice detection and speech, with all of it open source. The post says Marvin connects to ChatGPT and Entire, and it will be shown at the WeAreDevs booth 753 in San Jose.

AIUnsloth has passed 500 million model downloads on Hugging Face. Its Qwen3.8-27B GGUF is the most-downloaded model in Unsloth's history.

AIHugging Face has released a new JavaScript package, huggingface/lerobot, that lets developers read LeRobot datasets on the Hub directly in the browser without downloading them. The post says coding agents can use it to build custom dataset viewers quickly, with an example UI implemented in about 600 lines of JS.

AINVIDIA's Nemotron 3 Diarization is now available on ModelScope, labeling speakers and timestamps in streaming audio for up to eight speaker slots per conversation. The 99.2M-parameter model uses an end-to-end streaming architecture built on NVIDIA's Streaming Sortformer, running on Ampere, Hopper, and Blackwell GPUs via NeMo Speech C++. It is designed to pair with existing ASR systems such as Nemotron ASR, Parakeet, Canary, or Whisper to produce speaker-attributed transcripts.

AIA 25-second shampoo product demo, which many viewers believed was real footage, was AI-generated from a reusable UGC product-review template in the open-source Awesome Seedance repo. The repo contains 463 sourced cases, 264 cross-model retest runs, and 25 adaptable templates.

AIAlibaba's Qwen launched Qwen Intelligence with three mobile agents: a Mobile Planner Agent, a Mobile-Use Agent, and a Mobile Creative Agent. The post reports benchmark results including MobileWorld 82.1, MobileWorld-Real 92.2, and AndroidDaily 97.2, plus a 90% end-to-end success rate, and says the MobilePA-Bench, MobileWorld, MobileWorld-Real, and MobileWorld-Safety benchmarks are open.

Why it matters: The post names three mobile agents and their benchmark results, while also releasing the benchmark suite, so readers can check the claims against the reported figures.

AITeleOCR, a lightweight 1.2B vision-language model released under Apache 2.0, parses digital PDFs and warped phone photos without a separate dewarping model. It scores 96.87 overall on OmniDocBench v1.6, the highest among listed specialized VLMs, and ranks #1 in the ICDAR 2026 Sci-ImageMiner Challenge. It supports structured parsing of text, tables, formulas, layouts, and reading order, with synchronous or asynchronous vLLM inference.

AIXiaomi has released MiMo-V2.6 as an open model family under the MIT License, designed for large-scale reinforcement learning. MiMo-V2.6-Pro scores 46 on the Artificial Analysis Intelligence Index, with 71.9 on DeepSWE v1.1, 89.9 on Terminal-Bench 2.1, and 82.0 on OSWorld-Verified. The 1.02T-parameter MoE activates 42B parameters and supports text, image, video, and audio input with a 1M-token context.

Why it matters: The post links benchmark results, parameter scale, and a multi-agent RL training run, giving readers concrete figures to compare against other open models.

AIShanghai AI Lab and SJTU's LUMIA Lab released NCP-ArchPreview, an 8.9B open-weight language model under Apache 2.0. The model reportedly reaches OLMo-3-7B's final Stage 1 loss using 51.3% of the tokens from the 5.73T Dolma 3 corpus, a 1.95× convergence gain. Its concept module jointly predicts tokens and concepts, and domain adaptation updates only its 17M parameters while the token backbone stays frozen.

Why it matters: The post pairs an Apache 2.0 open-weight release with training-efficiency figures, showing how the concept module adapts to new domains with few trainable parameters.

AIinclusionAI open-sources the Ming-Image-0.1-Design family, two complementary 6B models for visual-design workflows, under an MIT License. Design generates complete UIs, dashboards, infographics, and posters up to 2048×2048 with native transparent RGBA output, and Layer decomposes flattened graphics into independently editable RGBA layers.

Why it matters: The post separates a design-generation model from a layer-decomposition model, letting users compare two distinct visual-design workflows under one MIT license.

AIUnsloth announced that Qwen-Image-2.1 FP8 and GGUF quantized versions should now run properly in Unsloth Desktop. The app supports both image generation and image editing with these quants. Further details are available on the Unsloth GitHub repository.

AIBoris Cherny used Opus 5.5 to formally verify the Claude Agent SDK with Lean, and short prompts produced 16 PRs fixing bugs and race conditions. He also combines Lean and TLA+ to find issues in data flow, concurrency, and state management, and says Claude is strong in both languages even though he does not know them well.

AIThe MiMo-V2.6 paper thread reports on the newest open model, which also streams its RL run, focusing on data and RL experimental results rather than architecture. The Pro model reportedly rose from 58.41 to 72.57 on DeepSWE after RL, with the top published DeepSWE score cited at 74.

AIGoogle Gemma says DiffusionGemma-Jev (djev) can now be deployed as a Jev API-compatible endpoint on Google Cloud Run with a single command. The post reports about 35-60 ms single-step latency and roughly 100-123 requests/sec at batch size 32, at about $3/hr that drops to $0 when idle.