AI Model2026
A 4B-parameter on-device general-purpose LLM for chat, writing, translation, coding and agentic workflows with native 1,000,000-token context. Uses a hybrid attention design to enable long-context efficiency, pretrained on ~20T tokens, and compatible with vLLM, llama.cpp, Ollama and LM Studio.
AI Dataset2026
Behavior-focused text corpus for LM pretraining, organized into seven Parquet-backed subsets (reasoning, planning, data-science, games, general, format-rewrites, other). Supports streaming, custom sampling, and large-scale dataset pipelines for research and model training.
AI Dataset2026
Provides multiple Parquet-backed subsets of code problem-solving data (direct answers, chain-of-thought reasoning, and task synthesis) that are streamable and prepared for language-model training and evaluation.
AI Dataset2026
Provides large-scale mathematical problem-solving, rewriting, and dialogue data organized into five Parquet-backed subsets for reasoning-oriented language-model training. Subsets support streaming access, Dataset Viewer inspection, and per-subset provenance metadata; licensed Apache 2.0.
AI Model2026
GGUF-quantized, refusal-removed build of Qwen3.8-Flash-Next for llama.cpp that provides multimodal (image+text), reasoning and tool-calling capabilities; released for security research and red-teaming under the Apache-2.0 license.
Analyzes on-policy self-distillation for language-model reasoning, diagnosing “collapse” and framing it as controlled by three levers: where token-level signals apply, what privileged information the teacher sees, and how teacher dynamics evolve.
A zero-data self-evolution framework that co-trains a Challenger, Solver, and Judge so LLMs can iteratively improve on both verifiable and unverifiable tasks without human labels. Uses role-asymmetry and subtask-amplification preference pairs to train the Judge and sustain improvement.
AI Model2026
A 770B-parameter Mixture-of-Experts instruct model from Tencent that natively supports 1,048,576-token contexts, Gated DSA attention, and speculative MTP decoding; open-sourced under Apache-2.0 with BF16 and FP8 weights for deployable inference.
Aozhe Wang, Zhengxi Lu +9Affiliation: Zhejiang University, Affiliation: Alibaba Group{waz,zhengxilu,syl}@zju.edu.cn [email protected]
A test-time method that adapts LLMs without labels by distilling rollouts that agree with majority pseudo-labels and penalizing disagreeing rollouts via grouped RL, improving robustness under frequent pseudo-label errors.
Provides a real-scale 3D Hong Kong sandbox to evaluate whether multimodal LLM agents can turn local street-view perception into sustained spatial action, supporting closed-loop first-person interaction, an interactive map, and controlled tests of grounding, long-range navigation, and robustness.
AI Dataset2026
Synthesizes 234K self-contained, high-difficulty scientific reasoning QA pairs by distilling research papers into compact 'reasoning skeletons'. Emphasizes mechanistic reasoning, hypothesis falsification, quantitative derivation and boundary calibration; built for SFT and reasoning evaluation.
AI Dataset2026
Provides 997 chain-of-thought cybersecurity reasoning records distilled from the Kimi K3 model, each with an explicit <think> trace and a technical resolution or structured tool invocation. Includes verified tool-call objects, diffs, cross-domain coverage, and token-level metadata for fine-tuning and evaluating reasoning models.