개발자

Compact foundation model tuned for efficient edge and on-device deployment. Arcee AI, 4.5B.

Purpose-built for semantic search, dense retrieval, and RAG. BAAI, 335M.

Multi-lingual, multi-granularity retrieval across dense, sparse, and multi-vector modes. BAAI, 568M.

Refreshed R1 checkpoint - mixture-of-experts reasoning with a fraction active per token. DeepSeek, 671B.

Mixture-of-experts reasoning powerhouse rivaling closed frontier models on math and code. DeepSeek, 671B.

R1 reasoning traces distilled into a Llama 70B backbone for math and code. DeepSeek, 70B.

Instruction-tuned text-diffusion Gemma 4 MoE that writes whole 256-token blocks per step, 4B active of 26B total. Google, 26B.

Distilled Whisper large-v3 — near-parity transcription at roughly six times the speed. Distil-Whisper, 756M.

Third-generation Falcon instruction model with a 32K context window. TII, 7B.

Flow-matching transformer for photorealistic text-to-image with strong prompt adherence. Black Forest Labs, 12B.

FLUX distilled to 4 steps — full quality, fraction of the compute. Black Forest Labs, 12B.

Largest Gemma 3 — 140-language reasoning, coding, long context. Google, 27B.

Instruction-tuned Gemma 4 for multilingual reasoning, coding, and long context. Google, 31B.

Open GPT-style model for deep reasoning in self-hosted deployments. GPT-OSS, 120B.

128K context, multilingual, tool-use ready, fully open weights. Meta, 70B.

Complex reasoning, coding, and agentic pipelines — no license restrictions. Meta, 70B.

128K context and multilingual base with a wide fine-tuning ecosystem. Meta, 8B.

Multilingual instruction following, tool use, and function calling. Meta, 8B.

Sub-gigabyte base model for on-device and embedded inference. Meta, 1B.

Instruction-tuned to run anywhere — on-device at 1B. Meta, 1B.

On-device reasoning with room for language understanding and light tool use. Meta, 3B.

Lightweight agent backbone for low-latency, resource-constrained deployments. Meta, 3B.

Stronger structured tasks, tool use, and reasoning than prior Llama generations. Meta, 70B.

Function-calling and instruction following for production and agentic use. Mistral AI, 7B.

Inverted residuals and linear bottlenecks — strong accuracy at near-zero inference cost. Google, 3.4M.

Text-to-video focused on motion quality and temporal coherence. Genmo, 10B.

Chinese, English, code, and structured output — 128K context base. Alibaba, 72B.

Instruction-tuned across Chinese, English, coding, and structured output at scale. Alibaba, 72B.

Toggleable chain-of-thought for on-demand deep reasoning. Alibaba, 32B.

Mid-scale Qwen3.6 for general reasoning, coding, and multilingual generation. Alibaba, 27B.

Reasoning depth on demand without a throughput penalty. Alibaba, 8B.

Compact multilingual embeddings for retrieval on constrained hardware. Alibaba, 0.6B.

Multilingual embeddings for retrieval and semantic similarity. Alibaba, 4B.

Higher-capacity multilingual embeddings for retrieval and reranking. Alibaba, 8B.

Hybrid thinking + vision — documents, charts, UI, and multi-image tasks. Alibaba, 32B.

Chain-of-thought with self-reflection for math, science, and logic. Alibaba, 32B.

Skip connections that made deep networks trainable — the image classification baseline. Microsoft Research, 25M.

Mix-transformer segmentation without positional encoding - accurate at low compute. NVIDIA, 3.8M.

Unified encoder-decoder for natural text-to-speech synthesis. Microsoft, 307M.

MMDiT architecture — strong text rendering and prompt control for image generation. Stability AI, 8B.

Dual-encoder base for high-resolution synthesis and fine-tuning pipelines. Stability AI, 6.6B.

SD-XL base driven image-to-image - edits or restyles an existing image from a prompt. Stability AI, 6.6B.

Image patches + self-attention, no convolutions — the original vision transformer. Google, 86M.

One-Shot Aggregation avoids DenseNet's redundant paths for better accuracy-per-FLOP. 11.2M.

Causal video transformer for text-to-video with strong motion coherence. Alibaba, 14B.

99 languages, 680K hours of training audio — built for robust speech recognition. OpenAI, 1.5B.

Anchor-free detector tuned for real-time inference at minimal parameter cost. Megvii, 0.9M.

Few-step distilled text-to-image generation. Tongyi-MAI.