MIT agent runtime. Pay model API only. Live Flash peak $0.44/$1.32 per 1M, not the old $0.14/$0.28 row.

DeepSeek small agent model: 284B/13B MoE, 1M context, MIT weights, $0.14/$0.28 per 1M tokens.

DeepSeek flagship 1.6T MoE with 49B active, 1M context, MIT weights. API $0.29/$1.20 cached $0.03.

DeepSeek-native AI coding agent for the terminal. A config- and plugin-driven harness in a single static Go binary, tuned around DeepSeek's prefix cache to k.

Task-driven workspace where a Chief agent splits one goal into todos, assigns them to agents by trade, and tracks progress, with 700+ models supported.

OpenAI-compatible AI API gateway: one base URL and key for 60+ models from GPT, Claude, Gemini, DeepSeek, Sora, and VEO, with routing, failover, and unified.

DeepSeek-R1 is a 671B-parameter reasoning model (37B active per token) trained with reinforcement learning, released in January 2025 under the MIT license.

DeepSeek never shipped a model named Coder-V2.5; DeepSeek-V2.5 (Sept 2024) merged V2 Chat and Coder-V2, and DeepSeek's API now leads with V4 models.

cogvlm-base-490-hf is an open CogVLM-17B vision-language checkpoint from Z.ai (formerly THUDM) with a visual expert module and Apache-2.0 code.

DeepSeek-V2.5, released September 5, 2024, merges DeepSeek-V2-Chat and Coder-V2-Instruct into one 236B open-weights model for chat and code.

DeepSeek V3 is a 671B-parameter Mixture-of-Experts model (37B active) trained on 14.8T tokens with a 128K context; DeepSeek's API now lists V4 models.

DeepSeek V4 is the anticipated next flagship DeepSeek model after V3, expected to improve reasoning, multimodal processing, and agent capabilities.

DeepSeek-VL-7B-Base is an open 7B vision-language model with a SigLIP-L and SAM-B encoder for 1024x1024 images, built on DeepSeek-LLM-7B-Base.

QwQ-32B-Preview is Qwen's experimental 32B reasoning model from November 2024, built on Qwen2.5-32B with 32,768 context under Apache 2.0; QwQ-32B followed.

DeepSeek's Apache-2.0 offline inference framework for DeepSeek V4.1 Flash, targeting large-scale serving on NVIDIA SM100 GPUs or Huawei Ascend 950 NPUs.

DeepSeek-V4.1-Flash is a MIT 552B multimodal MoE with 8B/16B activation, 1M context, and about 1/4 the KV cache of V4-Flash.

DeepSeek-V4-Flash-Vision-Exp is the first V4 multimodal experiment: MIT vision weights on Flash, 305B params, screenshots plus text agents.