AI Model2026
Early pretraining checkpoint of a compact multilingual causal LM aimed at low-memory deployment and Indic language support. Explores a Shared KV cache mode that can cut KV-cache memory by ~50% for inference; results are provisional (not a final, fully trained model).
AI Agent2026
An A‑share–specialized fork of TradingAgents that runs a seven‑analyst multi‑agent investment research pipeline for China stocks, integrating free A‑share data connectors and LLM providers. Key features: mootdx/東財 data integrations, A‑share trading rules (T+1, limits), Streamlit UI, and Apache‑2.0 license.
AI Dataset2026
25,000 chat-formatted synthetic SFT examples distilled to emulate the reasoning style and agentic behavior of Anthropic's Claude Mythos, focused on cybersecurity, advanced coding, mathematical reasoning, and long-horizon agent tasks. Includes metadata for targeted curriculum fine-tuning and is Apache-2.0 licensed.
AI Model2026
A trillion-parameter reasoning model aimed at long-horizon, multi-step agent workflows and tool collaboration. Offers adjustable Reasoning Effort modes (high, xhigh), async RL training (IcePop), and very long context (128K→256K) for complex production scenarios.
AI Model2026
Multimodal 35B scientific foundation model for image+text-to-text reasoning and conversational workflows. Uses task-scaling and full-chain training (pretraining → RL) to boost domain scientific abilities while keeping general multimodal reasoning and agent skills.
AI Model2026
A GGUF-format 9B model derived from Qwen3.5, fine-tuned for agentic coding, tool-calling, reasoning and vision-capable multimodal prompts. Optimized for local 8‑bit inference on 16GB-class machines; community experimental release for research use.
AI Model2026
Reasoning-enhanced 27B dense LLM fine-tuned from Qwen3.6-27B and released in GGUF format for image-text-to-text and long-context reasoning. Augmented with Trace Inversion reconstructed chains, three-stage SFT curriculum and MTP/vision support; community research release.
AI Model2026
Provides a 1-billion-parameter English pretrained language-model checkpoint that uses a dual-timescale Hierarchical Reasoning Model to increase effective compute depth. It's a PrefixLM pre-alignment checkpoint with composite-prefix modes for chain-of-thought style outputs; not instruction-tuned and requires downstream SFT/RL for assistant use.
AI Model2026
W4A4-quantized build of a 25B-parameter multimodal LLM that produces text from image+text inputs and supports conversational tool use. Trades very small quality differences for much lower GPU memory and latency so inference can run on smaller accelerators (vLLM support).
AI Model2026
A GGUF-format 9B LLM fine-tuned for code generation and agentic tool-calling that uses Multi-Token Prediction (MTP) and draft heads to increase throughput and long-range planning. Intended for local inference and research/experimental coding workflows; Apache‑2.0 license.
AI Dataset2026
Dataset of 5,000 reconstructed chain-of-thought samples produced by trace‑inversion from Claude‑opus‑4.7 summaries — packaged for SFT/DPO fine‑tuning. Key features: reconstructed CoT traces, multilingual prompts, gzip .jsonl format. Best used for reasoning distillation and model-level supervision; synthetic traces may need extra verification.
AI Dataset2026
Provides 9,000 reconstructed chain-of-thought (CoT) SFT examples produced by trace inversion from Claude Opus 4.6 outputs for fine-tuning reasoning-capable LLMs. Multilingual, packaged as .jsonl.gz and SFT/DPO-ready; verify numeric/code cases before training.