Uses Parallel Looped Transformers (PLT) to make loop count a practical knob for code models, finding two loops give the best test-time gains. Trains 7B models on 18T tokens and attributes saturation beyond two loops to a gain–cost tradeoff from positional mismatch.
AI Model2026
Provides FP8-quantized weights of GLM-5.2 — a 744B long-context LLM tuned for sustained 1M-token engineering, coding and agentic workflows; compatible with vLLM, Transformers, SGLang and Ascend NPU deployments.
Embodied AI2026
Provides a harness that lets language models control embodied manipulation via iterative perception–reasoning–action loops, semantic action abstractions, and multimodal observations. Demonstrates distilling capabilities into a 4B open-source model with under 2K simulated trajectories and shows sim-to-real generalization.
Proposes ZPPO, a distillation method that keeps the teacher inside prompts rather than injecting teacher gradients, using binary- and negative-candidate prompts plus a prompt replay buffer to recover learning signal on hard examples; shows gains for small Qwen3.5 students across 31 multimodal benchmarks.
AI Dataset2026
Provides 1.8M synthetic Belgian personas (1.2M records; 300k per language) in Dutch/French/German/English, grounded in Belgian census distributions to improve representativeness for LLM training and evaluation. Includes 23 persona and contextual fields, CC BY 4.0 license, produced with NeMo Data Designer.
AI Model2026
Provides GGUF-quantized GLM-5.2 builds for local text-generation with a solid 1M-token context, dynamic 1-/2-bit quant options, and Unsloth runtime integrations — targeted at long-horizon coding, reasoning and agent workflows. MIT licensed.
AI Dataset2026
Contains 603 coding and math prompt–response pairs produced by Claude Fable‑5 (generated 2026-06-10), provided as a JSONL subset for fine-tuning, evaluation, and behavior analysis. Responses are 'non-thinking' (no chain-of-thought); small, anonymized, and lacking an explicit license.
Evaluates multimodal LLMs' ability to reconstruct past observations and act in controllable non-Markov games. Introduces RNG-Bench with two games (Matching Pairs, 3D Maze), three controllable difficulty axes, a head-to-head duel protocol, and a Memory Gap metric to separate forgetting from action errors.
AI Dataset2026
Provides pre-parsed arXiv LaTeX source files aligned with official metadata as ready-to-query Parquet rows. Bundles each paper's .tex/.bib/.sty etc. into a single readable tree, updates monthly, and simplifies large-scale access for LLM pretraining, document understanding, and citation analysis while requiring adherence to original arXiv licenses.
AI Agent2026
Self‑hosted A‑share quantitative workbench for screening, monitoring, backtesting and stock-level analysis using TickFlow data; supports 18 Polars strategies, vectorbt backtesting, realtime rule-based alerts, pluginable data sources and optional LLM-driven strategy generation and stock analysis.
AI Model2026
A local, quantized Gemma 4 12B fine-tune packaged as GGUF quants that runs on ~4.5 GB VRAM. Optimized for coding and multi-step agentic tool use (read→reason→act→verify), ships multiple quant sizes (Q3_K_M–Q8_0) and MTP draft support; English-centric with trade-offs versus generalist models.
AI Dataset2026
Provides a deduplicated 2.0M-row corpus of FABLE.5 / Mythos agent traces with row-level provenance and session-limit rows removed. Includes canonical Parquet and gzip JSONL exports, SHA256 row hashes, and provenance fields for tracing first-source datasets.