Discover the Best AI Resources

Step 3.7 Flash

Processes images and text to produce structured, reasoning-rich text outputs for high-throughput agentic workflows. Sparse MoE design (198B total, ~11B active per token), 256k context window and selectable reasoning levels—optimized for single-pass parsing, verification, and multi-step automation.

StreamAudio-2M

Large streaming-audio dataset for training and evaluating audio-LLMs and audio agents. About 2.28M clips grouped into multi-turn “streams” across six task subsets (ASR, speech translation, audio understanding, voice chat, proactive response, environment-aware); audio shipped as tar shards.

MOSS-TTS-v1.5

Generates multilingual text-to-speech with zero-shot voice cloning, token-level duration control, and inline pause markers. v1.5 improves multilingual fidelity (with language tags), cloning stability, and long-reference handling—suitable for research and production TTS pipelines.

Keye-VL-2.0-30B-A3B

Performs hour-scale video understanding and fine-grained temporal localization while exposing agent-style multimodal tool/code/search abilities. Built on a sparse-attention long-context architecture (DSA) and a specialized inference stack—best used in GPU-backed research or production evaluation.

Mellum2 Thinking

Generates text with explicit chain-of-thought traces for multi-step reasoning and math-heavy tasks, emitting reasoning inside <think>...</think> blocks. Uses a Mixture-of-Experts design and 131k token context for long, verifiable workflows—best when you need inspectable reasoning.

LocateAnything-3B

Performs fast, high-quality vision–language grounding: given an image plus a natural-language prompt it returns bounding boxes or points for referred objects. Uses Parallel Box Decoding for parallel coordinate prediction (higher throughput) and targets research/non-commercial use.

Gajae-Code

Runs an external, reviewable coding-agent harness that turns intent into repeatable software work: clarifies requirements, builds reviewed plans, executes in tmux-backed sessions, and collects durable verification. Ships Telegram/Discord delivery, a research REPL, and optional desktop-control tools; beta-stage.

MeasL-Bench

Benchmark for evaluating vision–language models on measurement-grounded inputs vs. RGB, emphasizing low-light, HDR, and visibility-sensitive evidence recovery. Contains 2,183 paired test examples with local image assets for controlled RAW↔RGB comparisons.

PaddleOCR-VL-1.6

Performs image-to-text document parsing and OCR for complex elements (tables, formulas, charts, seals), with multilingual support (en/zh). It uses region-aware data optimization and progressive post-training to improve weak-region supervision and is plug-and-play compatible with PaddleOCR-VL-1.5.

Skill0.5: Joint Skill Internalization and Utilization for Out-of-Distribution Generalization in Agentic Reinforcement Learning

Combines internalizing general skills with task-specific skill utilization via a difficulty-aware router to improve in-distribution and out-of-distribution performance for agentic RL. Uses privileged distillation for hard tasks and diagnostic probing for easy tasks; evaluated on ALFWorld and WebShop.

nvidia/Qwen3.6-35B-A3B-NVFP4

Quantized NVFP4 build of the Qwen3.6-35B MoE language model, optimized with NVIDIA Model Optimizer to cut model size and GPU memory by ~3.06× for inference. Designed for vLLM and NVIDIA GPU deployments (Hopper/Blackwell).

A Matter of TASTE: Improving Coverage and Difficulty of Agent Benchmarks

Proposes TASTE, an automatic pipeline that synthesizes challenging agent benchmark tasks by sampling and evolving valid tool-sequence patterns; uses an adaptive contrastive n-gram model and LLM validity judgments to produce τ^c-Bench with broader tool-use coverage and higher difficulty.