AI Model2026
Mixture-of-Experts LLM tuned for mathematical and coding reasoning, with ~760M active / 8.4B total parameters and post-training for improved stepwise reasoning. Optimized for inference efficiency (vLLM/transformers forks) so it can run in computation-constrained or local deployments; Apache-2.0 licensed.
AI Dataset2026
Agentic coding evaluation dataset containing real-world, multi-step developer tasks and raw model responses across 20+ programming languages. Emphasizes challenging, persona-driven prompts for benchmarking and fine-tuning; users should filter and audit outputs before training.
AI Model2026
Merges Unsloth UD XL quantized GGUF of Qwen3.6-27B with compact Q8_0 MTP heads to enable multi-token (speculative) decoding on llama.cpp builds that support MTP; aimed at image-text-to-text usage with reduced MTP overhead.
AI Infra2026
High-throughput LLM inference engine for agentic workloads, combining a local‑SPMD static compiler for parallelism, a C++ scheduler with a Python execution plane and type‑safe KV‑cache reuse, pluggable high-performance kernels (including an MLA implementation), and a low‑overhead AsyncLLM entrypoint for production GPU inference.
AI Dataset2026
A retrieval benchmark suite focused on “oblique queries,” where relevance depends on latent attributes rather than surface keywords. Includes five tasks with large corpora, qrels (and pooled judgments), and task-specific constraints for evaluating embedding-based retrievers and reasoning-augmented retrieval.
Preview of an MoE model family (V4-Pro: 1.6T params, 49B active; V4-Flash: 284B, 13B active) built for 1M-token contexts. A hybrid attention design cuts single-token inference FLOPs to 27% and KV cache to 10% versus V3.2 at million-token length.
AI Model2026
A reasoning-enhanced Mixture-of-Experts (MoE) LLM fine-tuned for multimodal image-text-to-text tasks and long-context reasoning; built on Qwen3.6-35B-A3B with LoRA and released as an experimental GGUF community model.
AI Dataset2026
Provides JSON traces from a Codex-driven swebenchpro agentic benchmark, including per-call token counts, cache hit rates, timing, and per-trial outcomes. Useful for research into LLM caching, long-context workloads, and agent evaluation. MIT-licensed and compact.
AI Agent Papers2026
Trains one model to delegate work recursively to copies of itself. Joint reinforcement learning teaches task splitting and communication, improving harder-task generalization while trading extra computation for parallel execution.
AI Dataset2026
Provides 1,781 OpenTelemetry execution traces of LLM-powered agents across six benchmarks, including full conversations, token usage, timing, tool calls and model metadata—useful for performance analysis, agent-behavior research, and inference debugging.
AI Agent2026
Trains reusable natural-language 'skills' for frozen LLM agents by optimizing the skill document in text-space — using trajectory-driven edits, validation-gated updates, and deployable best_skill.md artifacts. Multi-backend, zero inference-time cost at deployment, designed for iterative, validation-led skill improvement.
AI Dataset2026
Contains 4,006 newline-delimited JSONL agent-session traces recording assistant responses and tool calls from deepseek/deepseek-v4-pro — includes a training-ready tools schema snapshot and helpers for conversion to SFT/distillation workflows.