AI Model2026
A dense ~9B reasoning LLM optimized for agentic coding and tool-calling that emits explicit chain-of-thought (<think>) blocks and well-formed tool calls. Designed to run on a single 80GB GPU (~19GB bf16), uses self-scaffolding RL and exposes an OpenAI-compatible API.
AI Audio2026
Generates streaming, low‑latency neural speech for real‑time dialogue by autoregressively producing audio frames as text arrives; joint text–speech training preserves natural prosody. Optimized for vLLM streaming (~50 ms first chunk), supports short‑clip voice cloning and four languages.
AI Model2026
Simulates agentic environments and predicts next environment states from actions and interaction history using a language-based world model across seven domains. Trained via a CPT→SFT→RL pipeline with an MoE architecture and very long context; intended for environment simulation and agent research.
AI Dataset2026
Provides 2,170 reference-grounded evaluation samples across seven agent domains (MCP, Search, Terminal, SWE, Android, Web, OS) to score language world models on Format, Factuality, Consistency, Realism and Quality. Includes per-domain JSONL files, judge prompts and an evaluation script for reproducible scoring.
AI Model2026
NVFP4-quantized variant of Qwen3.6-27B that reduces parameter bits from 16 to 4, cutting disk and GPU memory requirements by ~2.5× while keeping comparable benchmark accuracy; ready for vLLM-based inference on NVIDIA hardware and supports long, multimodal contexts.
AI Model2026
Provides an open-source Mixture-of-Experts coding LLM (397B) optimized for agentic, tool-enabled coding workflows with a 262,144-token context window, OpenAI-compatible API, serving recipes (vLLM/SGLang), and published coding-benchmark results.
Proposes Monotonic Inference Policy Improvement (MIPI) and a two-step Monotonic Inference Policy Update (MIPU) to address training–inference probability mismatch in LLM reinforcement learning by constructing sampler-referenced candidate updates and accepting synchronized updates using an inference-gap proxy; shows improved reasoning accuracy and stability under FP8-quantized rollouts.
Treats hybrid layer selection as a budget-constrained subset optimization and introduces FlashMorph: a pipeline that equips each transformer layer with a linear-attention branch, jointly optimizes layerwise gates on synthetic long-context retrieval data, then discretizes, distills, and finetunes—achieving strong long-context recall using only 20M selection tokens.
AI Model2026
Thinking-off fine-tune for coding-agent workflows that prioritizes fast next-step decisions, lower token usage and stable multi-turn tool calling. Highlights: MoE 35B base, MTP speculative decoding, SWE-bench 62.4% (300 cases). Best for local agent loops and automated debug cycles; requires disciplined harnessing and schema consistency.
AI Model2026
Provides GGUF/llama.cpp quantized variants of Qwen3.6-27B for local multimodal inference, tuned via online RL to cut average 'thinking' tokens by ≈50% while preserving answer quality; offers Q4_K_M/Q8_0/f16 builds and a separate mmproj for vision input.
Xingyu Zheng, Xianglong Liu +5State Key Laboratory of Complex & Critical Software Environment, Beihang University, School of Computer Science and Engineering, Beihang University +6
Accelerates text-to-image diffusion for pretrained flow-matching models using a staged low-to-high-resolution pipeline: fast low-res sampling, pixel-space GAN super-resolution, light latent noising, and short high-res refinement — >10× end-to-end speedups without retraining.
Compiles natural-language function specifications into compact, locally-executable neural programs (PAW) that run on a small frozen interpreter; a 4B compiler emits LoRA adapters for a 0.6B runtime to provide offline, low-memory fuzzy text functions.