Model Marketplace

DeepSeek-V4-Flash is a lightweight MoE text model built for online workloads at scale: 284B total parameters with only about 13B activated per inference. It keeps long-context comprehension while optimizing the reasoning path for everyday chat, retrieval-augmented generation, content production, customer support, coding assistance and long-running tasks. Its unit price is markedly lower than comparable models, making it a good fit for high-QPS, low-latency and cost-sensitive integrations.

VendorDeepSeek

DeepSeek-V4-Pro is a flagship MoE model for business-critical paths: 16T total parameters, 498B activated per inference, with native million-token context. It is optimized for hard reasoning and high-quality training data, with particular strength in mathematical logic, complex reasoning, professional code generation and long-document analysis — suited to high-throughput workloads and multi-step automation pipelines as a core engine on the critical path.

VendorDeepSeek

GLM-5.2 is the flagship base model of the GLM-5 series, covering text generation, deep reasoning and code engineering. Aiming at "staying usable across long tasks", we redesigned the context-compression path and the goal-consistency mechanism so the model reliably retains key constraints across multi-day engineering collaboration, and keeps aligning with user intent through long-chain tasks without context drift. Combined with tiered thinking effort, function calling, MCP tools and structured output, it can drive project-scale software development and complex automation pipelines.

VendorZhipu AI

GLM-5.3 is the strongest open-source model for coding released by Zhipu. It shares the base model and parameter scale of GLM-5.2 (743B total parameters), with every capability gain coming from extreme post-training scaling. It breaks new ground on complex software engineering, terminal operations and long-horizon agent tasks, with coding capability 50% higher than GLM-5.2; it also fills in cybersecurity, performing strongly on defensive security tasks such as white-box code review and vulnerability discovery. Combined with a million-token context, tiered thinking effort, function calling and MCP tools, it can drive project-scale software development, security audits and complex automation pipelines.

VendorZhipu AI

Kimi K2.7-Code is a Moonshot AI model fine-tuned specifically for coding, scoring at the top of mainstream code benchmarks including HumanEval, MBPP and BigCodeBench. It supports 70+ programming languages and the major IDE protocols, handling completion, unit tests, bug fixes and cross-file refactoring in one flow — a capable assistant for day-to-day engineering and automated code workflows.

VendorMoonshot AI

Kimi K3 is Moonshot AI's flagship multimodal model, built around a 1M-token context window and tiered deep reasoning. It keeps hold of the goal through long-document reading, cross-chapter references and multi-turn tool calls; combined with function calling, MCP tools and structured output, it can drive enterprise knowledge-base Q&A and complex automation pipelines.

VendorMoonshot AI

GLM-5.3-Flash is the first natively multimodal model in the GLM-5 series. Built on a 320B-A18B hybrid of sparse and linear attention, it is the world's first open frontier model to adopt that architecture. It scores 57 on the Artificial Analysis Intelligence Index, matching Claude Opus 4.8 and surpassing GLM-5.2 across the board. Vision is woven natively into the coding loop, so the model watches interfaces, rendered output and interaction feedback and keeps improving, working across code, browser and GUI to finish a task. It also reaches into Office work, financial research and professional documents, decomposing goals on its own, calling tools and delivering finished PPTX, PDF, DOCX and XLSX files. At one tenth the price of GLM-5.3, frontier intelligence no longer has to be rationed.

VendorZhipu AI

DeepSeek-V4.1-Flash is the smallest model in DeepSeek's new architecture family, with native multimodal visual understanding. It is a 552B-parameter MoE model built on the new Causal-Encoder-Decoder architecture, with asymmetric input and output: only 8B parameters are activated for input and 16B for output, at a cost well below other known models of the same size. The new architecture sharply compresses the KV cache, cutting HBM requirements to 1/4 and SSD requirements to 1/8 of the previous generation, which lowers the cost of agent workloads where cache-hit charges make up a large share. The model supports a 1M context window and up to 384K output tokens; thinking mode is on by default, with three effort levels (low, high and max) to match task complexity.

VendorDeepSeek