τ₀-VLA

TurboVLA is a compact vision-language-action model and official implementation for robotic manipulation, developed by researchers at Huazhong University of Science and Technology and Huawei Technologies. Instead of routing visual input through a large language model, it independently encodes visual observations and language instructions, exchanges information through lightweight bidirectional vision-language interaction, and uses a compact decoder to predict continuous action chunks. The repository includes training and evaluation code for LIBERO and RoboTwin, with model checkpoints released separately. The project reports 32 Hz inference on an RTX 4090 with less than 1 GB of VRAM; on LIBERO, it reports a 97.7% average success rate, 31.2 ms inference latency, and 0.2 billion parameters.

LightNav-0 is an open-source generalist embodied-navigation model from the Light Origins Team. It uses a pretrained Qwen3-VL backbone with a single egocentric RGB stream and natural-language instructions to control humanoid, quadruped, wheeled, and aerial robots, transferring across tasks, embodiments, and scenes without per-task or per-benchmark fine-tuning. The model extends the vocabulary with dual-channel pointing tokens and residual vector-quantized action tokens rather than adding navigation-specific prediction heads. At each step, it emits an affordance point representing a feasible local direction or waypoint and an object point representing the goal, followed by three action tokens that decode to ten future SE(2) waypoints for an embodiment-specific low-level controller. Its temporally aware history compressor samples older frames less frequently, pools them more coarsely, and preserves ordering with timestamp tokens while bounding the visual context. The repository includes a released checkpoint and action decoder, command-line prediction, a vLLM-based server with WebSocket streaming, a MuJoCo simulation demo, evaluation harnesses, and a ROS 2 deployment stack with adapters for the Unitree Go2 and LimX TRON 1. The project is released under the Apache License 2.0; the included EVT-Bench is separately licensed CC BY-NC-SA 4.0.

vLLM is an open-source library and inference-serving engine for deploying large language models on GPUs and other hardware accelerators. Originally developed in UC Berkeley's Sky Computing Lab, it is maintained by a broad community and provides an OpenAI-compatible API server, as well as Anthropic Messages API and gRPC support. Its serving stack manages attention key-value memory with PagedAttention, continuous batching, chunked prefill, prefix caching, streaming generation, and disaggregated prefill, decode, and encode. It also supports speculative decoding, quantization, optimized attention and GEMM/MoE kernels, automatic kernel generation and graph-level transformations, structured outputs, tool calling, multiple decoding algorithms, and distributed inference through tensor, pipeline, data, expert, and context parallelism. The project integrates with Hugging Face models and supports decoder-only, mixture-of-experts, hybrid state-space, multimodal, embedding, retrieval, reward, and classification models. It can run across NVIDIA, AMD, Intel, and other supported accelerators, as well as x86, ARM, and PowerPC CPUs, and can be installed with uv or pip or built from source.

LingBot-World 2.0, also called LingBot-World-Infinity, is an AI world-modeling system from the Robbyant Team for generating interactive video environments. Its causal pretraining paradigm is designed for an unbounded interaction horizon, while a distilled real-time variant is claimed to drive 720p video streams at 60 frames per second. The system supports diverse actions, including attacking, archery, spell-casting, and shooting, along with text-driven events. Its agentic harness uses a pilot agent to plan and execute character behavior and a director agent to synthesize new environmental elements as a scene progresses. The causal inference implementation processes video frames chunk by chunk with KV caching rather than processing the entire sequence at once. The release provides inference code and model weights, including a 14B causal-fast model, with a codebase built on Wan2.2; models are distributed through Hugging Face and ModelScope, and the real-time system can be tried through Reactor on the web or LingGuang on mobile. The project is licensed under CC BY-NC-SA 4.0 for non-commercial use. It states that deployment code will not be released and points to SGLang or flashdreams for self-deployment references.

TrueForge is an open-source, vendor-neutral agent harness developed by TrueFoundry. It runs the agent execution loop, including model calls, MCP tool use, skills, sandboxed code and file execution, approvals, context management, and session state. Agents can use configured model providers, remote MCP servers, git-backed SKILL.md instruction packs, and on-demand sandboxes; context features include subagents, deferred tool loading, Code Mode, large-result offloading, and compaction. TrueForge exposes agents through a bundled chat UI, an HTTP API with a TypeScript SDK, and an embeddable UI SDK. It supports local mode with SQLite and hosted deployments using Postgres and Redis through Docker Compose or Helm; the repository warns that local mode is intended for personal use on localhost rather than production or internet-facing deployment.