Tag
Explore by tags
Open-weight Mixture-of-Experts LLM with 671B total parameters but 37B activated per token, trained on 14.8T tokens for 2.788M H800 GPU-hours. Matches leading closed models at a fraction of typical training cost via FP8 and architectural tricks.
Shows that LLM reasoning can be incentivized through pure reinforcement learning, with no human-annotated reasoning traces. Self-reflection, verification, and strategy-switching emerge on their own, and the patterns transfer to distill smaller models.
High-performance CUDA tensor-core GEMM kernel library for LLM workloads: supports FP8/FP4/BF16, fused Mega MoE and MQA scoring, and runtime JIT-compiled kernels. Targets NVIDIA SM90/SM100 and PyTorch—for teams working on low-level GPU kernel optimization.
Provides high-throughput, low-latency GPU communication kernels for Mixture-of-Experts (MoE) and expert-parallel workloads, with NVLink↔RDMA-aware forwarding, FP8/BF16 support, and low-latency RDMA hooks for inference decoding.
Optimized MLA (Multi-head Latent Attention) decoding kernels powering DeepSeek-V3/V3.2 inference on Hopper and Blackwell GPUs. Dense decoding reaches ~3000 GB/s and 660 TFLOPS on H800; the sparse path stores the KV cache in FP8.
Orchestrates a lead agent, isolated parallel sub-agents, long-term memory, and sandboxes for long-horizon tasks — minutes to hours of deep research, coding, and content creation. LangChain/LangGraph-based with extensible skills; v2 is a full rewrite.
An agentic framework that analyzes, plans, and executes multi-step video understanding and editing workflows using multimodal LLM-driven agents—features intent decomposition, graph-based workflow orchestration, and automated shot planning for long-form video tasks.
Enhances web AI chat interfaces (Gemini, AI Studio, Claude, ChatGPT) with folder-based conversation organization, timeline navigation, prompt vault, cloud sync, chat export, UI fixes and a plugin engine. Runs as a browser extension and includes cross-site prompt management and Mermaid rendering.
An open large language model pairing DeepSeek Sparse Attention (DSA) for cheaper long-context inference with a scaled RL pipeline. Authors claim parity with GPT-5, with a high-compute Speciale variant surpassing it and rivaling Gemini-3.0-Pro on reasoning.
Provides a conditional memory module that performs O(1) N‑gram lookups and fuses static embeddings into transformer hidden states — enables offloading large embedding tables to host memory with minimal inference overhead.
Terminal-native coding agent that streams reasoning blocks, makes controlled edits to local workspaces behind approval gates, and includes an auto mode that chooses model and thinking level per turn — designed for in-terminal code review, debugging, and automation workflows.
