Alternatives to DeepSeek-V2
Compare DeepSeek-V2 alternatives for your business or organization using the curated list below. SourceForge ranks the best alternatives to DeepSeek-V2 in 2026. Compare features, ratings, user reviews, pricing, and more from DeepSeek-V2 competitors and alternatives in order to make an informed decision for your business.
-
1
Kimi K3.1 Pro
Moonshot AI
Kimi K3.1 Pro is a multimodal large language model developed by Moonshot AI for advanced reasoning, software development, 3D modeling, financial analysis, and autonomous AI agent workflows. The model uses a mixture-of-experts architecture with approximately 2.8 trillion total parameters and 104 billion active parameters, supporting a context window of one million tokens. Its coding capabilities include debugging existing applications, generating complete software projects, developing interactive websites, and creating applications with game engines. Kimi K3.1 Pro can analyze images, interpret spatial relationships, generate CAD scripts, and create three-dimensional assets from text descriptions or visual references. Its reasoning and tool-use capabilities support financial modeling, investment research, document analysis, and multistep business workflows. The model is available through Kimi's web and mobile applications, Kimi Work, Kimi Code, and the Kimi API, with a Highspeed variant.Starting Price: $2 per 1M tokens -
2
DeepSeek-V4-Pro
DeepSeek
DeepSeek-V4-Pro is a large-scale Mixture-of-Experts (MoE) language model designed for advanced reasoning, coding, and long-context understanding. It features 1.6 trillion total parameters with 49 billion activated parameters, enabling high performance while maintaining efficiency. The model supports an exceptionally large context window of up to one million tokens, allowing it to process extensive documents and workflows. It uses a hybrid attention architecture to optimize long-context performance and reduce computational cost. DeepSeek-V4-Pro is trained on over 32 trillion tokens, improving its knowledge and reasoning capabilities. It also includes advanced optimization techniques for stability and faster convergence during training. The model supports multiple reasoning modes, allowing users to balance speed and accuracy based on their needs. Overall, it provides a powerful open-source solution for complex AI tasks and large-scale applications.Starting Price: $0.435 per 1M tokens (input) -
3
GLM-5.3-Flash
Z.ai
GLM-5.3-Flash is Z.ai’s natively multimodal model in the GLM-5 series (previously previewed as Ox Alpha), designed to deliver strong coding, agentic, visual, and knowledge-work performance at relatively low inference cost. It uses 320 billion total parameters with 18 billion active parameters, along with a hybrid architecture that combines sparse and linear attention to reduce the cost of long-context processing. The model supports context lengths of up to one million tokens and was trained on a 30-trillion-token multimodal corpus. GLM-5.3-Flash can reason across text, images, documents, interfaces, dashboards, and other visual information while using that feedback to refine its own outputs. Z.ai reports substantial gains over GLM-5.2 on coding and agentic benchmarks, including DeepSWE and AutomationBench, while approaching higher-cost frontier models on several evaluations.Starting Price: $0.15 per 1M tokens (input) -
4
Nemotron 3 Ultra
NVIDIA
Nemotron 3 Nano is a compact, open large language model in NVIDIA’s Nemotron 3 family, designed for efficient agentic reasoning, conversational AI, and coding tasks. It uses a hybrid Mixture-of-Experts Mamba-Transformer architecture that activates only a small subset of parameters per token, enabling low-latency inference while maintaining strong accuracy and reasoning performance. It has approximately 31.6 billion total parameters with around 3.2 billion active (3.6 billion including embeddings), allowing it to achieve higher accuracy than previous Nemotron 2 Nano while using less computation per forward pass. Nemotron 3 Nano supports long-context processing of up to one million tokens, enabling it to handle large documents, multi-step workflows, and extended reasoning chains in a single pass. It is designed for high-throughput, real-time execution, excelling in multi-turn conversations, tool calling, and agent-based workflows where tasks require planning, reasoning, and more. -
5
Inkling
Thinking Machines Lab
Inkling is an open-weights multimodal AI model from Thinking Machines designed as a customizable foundation model for developers, researchers, and enterprises. The model is a Mixture-of-Experts transformer with 975 billion total parameters, 41 billion active parameters, and support for context windows up to 1 million tokens. Inkling was trained from scratch on text, images, audio, and video, giving it native capabilities across reasoning, coding, agentic tool use, vision, audio, factuality, and instruction following. It is built with controllable thinking effort so users can balance performance, latency, and token efficiency for different workloads. The model is available for fine-tuning on Tinker, with playground access, API availability through ecosystem partners, and full weights published on Hugging Face. Built for customization, Inkling gives teams an open-weights base model for building domain-specific AI systems, multimodal agents, coding workflows, research tools, and more.Starting Price: Free -
6
DeepSeek-V4
DeepSeek
DeepSeek-V4 is a next-generation open-source language model designed for high-performance reasoning, coding, and long-context intelligence. It introduces a powerful architecture with up to one million token context length, enabling seamless handling of large datasets and complex multi-step workflows. The model comes in two variants: DeepSeek-V4-Pro for maximum performance and DeepSeek-V4-Flash for efficiency and speed. DeepSeek-V4-Pro features 1.6 trillion total parameters with 49 billion activated, delivering near state-of-the-art performance comparable to leading closed-source models. It excels in agentic coding, mathematical reasoning, and world knowledge tasks. The model integrates advanced attention mechanisms, including token-wise compression and sparse attention, significantly reducing compute and memory costs. It is also optimized for AI agents, supporting tool use and multi-step workflows.Starting Price: Free -
7
DeepSeek-V4.1-Flash-1017
DeepSeek
DeepSeek-V4.1-Flash-1017 is an open-weight multimodal artificial intelligence model developed by DeepSeek for advanced reasoning, software engineering, visual understanding, and autonomous AI agent workflows. Released on October 17, 2026, it builds on DeepSeek-V4.1-Flash with improvements in coding performance, agentic task execution, and reasoning efficiency. The model supports text and image inputs, tool calling, configurable reasoning effort, and multistep problem-solving. Its one-million-token context window enables developers to process lengthy documents, extensive codebases, and complex sequences of instructions. DeepSeek-V4.1-Flash-1017 also supports end-to-end software development, including application generation, debugging, automated testing, and interactive project creation. The model is available through the DeepSeek API and downloadable weights under the MIT License, supporting both hosted and self-managed deployment.Starting Price: Free -
8
DeepSeek-V4-Flash
DeepSeek
DeepSeek-V4-Flash is a high-efficiency Mixture-of-Experts (MoE) language model designed for fast, scalable reasoning and text generation. It features 284 billion total parameters with 13 billion activated parameters, delivering strong performance while optimizing computational cost. The model supports an extensive context window of up to one million tokens, enabling it to process large documents and complex workflows with ease. Its hybrid attention architecture enhances long-context efficiency by reducing memory and compute requirements. Trained on over 32 trillion tokens, DeepSeek-V4-Flash demonstrates solid capabilities across knowledge, reasoning, and coding tasks. It is designed for scenarios where speed and efficiency are critical, offering a balance between performance and resource usage. The model also supports multiple reasoning modes, allowing users to adjust between faster outputs and deeper analysis.Starting Price: $0.14 per 1M tokens (input) -
9
DeepSeek R1
DeepSeek
DeepSeek-R1 is an advanced open-source reasoning model developed by DeepSeek, designed to rival OpenAI's Model o1. Accessible via web, app, and API, it excels in complex tasks such as mathematics and coding, demonstrating superior performance on benchmarks like the American Invitational Mathematics Examination (AIME) and MATH. DeepSeek-R1 employs a mixture of experts (MoE) architecture with 671 billion total parameters, activating 37 billion parameters per token, enabling efficient and accurate reasoning capabilities. This model is part of DeepSeek's commitment to advancing artificial general intelligence (AGI) through open-source innovation.Starting Price: Free -
10
DeepSeek-V3.2
DeepSeek
DeepSeek-V3.2 is a next-generation open large language model designed for efficient reasoning, complex problem solving, and advanced agentic behavior. It introduces DeepSeek Sparse Attention (DSA), a long-context attention mechanism that dramatically reduces computation while preserving performance. The model is trained with a scalable reinforcement learning framework, allowing it to achieve results competitive with GPT-5 and even surpass it in its Speciale variant. DeepSeek-V3.2 also includes a large-scale agent task synthesis pipeline that generates structured reasoning and tool-use demonstrations for post-training. The model features an updated chat template with new tool-calling logic and the optional developer role for agent workflows. With gold-medal performance in the IMO and IOI 2025 competitions, DeepSeek-V3.2 demonstrates elite reasoning capabilities for both research and applied AI scenarios.Starting Price: Free -
11
DeepSeek-V3.2-Exp
DeepSeek
Introducing DeepSeek-V3.2-Exp, our latest experimental model built on V3.1-Terminus, debuting DeepSeek Sparse Attention (DSA) for faster and more efficient inference and training on long contexts. DSA enables fine-grained sparse attention with minimal loss in output quality, boosting performance for long-context tasks while reducing compute costs. Benchmarks indicate that V3.2-Exp performs on par with V3.1-Terminus despite these efficiency gains. The model is now live across app, web, and API. Alongside this, the DeepSeek API prices have been cut by over 50% immediately to make access more affordable. For a transitional period, users can still access V3.1-Terminus via a temporary API endpoint until October 15, 2025. DeepSeek welcomes feedback on DSA via its feedback portal. In conjunction with the release, DeepSeek-V3.2-Exp has been open-sourced: the model weights and supporting technology (including key GPU kernels in TileLang and CUDA) are available on Hugging Face.Starting Price: Free -
12
DeepSeek-V3.1-Terminus
DeepSeek
DeepSeek has released DeepSeek-V3.1-Terminus, which enhances the V3.1 architecture by incorporating user feedback to improve output stability, consistency, and agent performance. It notably reduces instances of mixed Chinese/English character output and unintended garbled characters, resulting in cleaner, more consistent language generation. The update upgrades both the code agent and search agent subsystems to yield stronger, more reliable performance across benchmarks. DeepSeek-V3.1-Terminus is also available as an open source model, and its weights are published on Hugging Face. The model structure remains the same as DeepSeek-V3, ensuring compatibility with existing deployment methods, with updated inference demos provided for community use. While trained at a scale of 685B parameters, the model includes FP8, BF16, and F32 tensor formats, offering flexibility across environments.Starting Price: Free -
13
GLM-4.6
Z.ai
GLM-4.6 advances upon its predecessor with stronger reasoning, coding, and agentic capabilities: it demonstrates clear improvements in inferential performance, supports tool use during inference, and more effectively integrates into agent frameworks. In benchmark tests spanning reasoning, coding, and agents, GLM-4.6 outperforms GLM-4.5 and shows competitive strength against models such as DeepSeek-V3.2-Exp and Claude Sonnet 4, though it still trails Claude Sonnet 4.5 in pure coding performance. In real-world tests using an extended “CC-Bench” suite across front-end development, tool building, data analysis, and algorithmic tasks, GLM-4.6 beats GLM-4.5 and approaches parity with Claude Sonnet 4, winning ~48.6% of head-to-head comparisons, while also achieving ~15% better token efficiency. GLM-4.6 is available via the Z.ai API, and developers can integrate it as an LLM backend or agent core using the platform’s API.Starting Price: Free -
14
DeepSeek-V3.2-Speciale
DeepSeek
DeepSeek-V3.2-Speciale is a high-compute variant of the DeepSeek-V3.2 model, created specifically for deep reasoning and advanced problem-solving tasks. It builds on DeepSeek Sparse Attention (DSA), a custom long-context attention mechanism that reduces computational overhead while preserving high performance. Through a large-scale reinforcement learning framework and extensive post-training compute, the Speciale variant surpasses GPT-5 on reasoning benchmarks and matches the capabilities of Gemini-3.0-Pro. The model achieved gold-medal performance in the International Mathematical Olympiad (IMO) 2025 and International Olympiad in Informatics (IOI) 2025. DeepSeek-V3.2-Speciale does not support tool-calling, making it purely optimized for uninterrupted reasoning and analytical accuracy. Released under the MIT license, it provides researchers and developers an open, state-of-the-art model focused entirely on high-precision reasoning.Starting Price: Free -
15
DeepSeek-Coder-V2
DeepSeek
DeepSeek-Coder-V2 is an open source code language model designed to excel in programming and mathematical reasoning tasks. It features a Mixture-of-Experts (MoE) architecture with 236 billion total parameters and 21 billion activated parameters per token, enabling efficient processing and high performance. The model was trained on an extensive dataset of 6 trillion tokens, enhancing its capabilities in code generation and mathematical problem-solving. DeepSeek-Coder-V2 supports over 300 programming languages and has demonstrated superior performance on benchmarks such surpassing other models. It is available in multiple variants, including DeepSeek-Coder-V2-Instruct, optimized for instruction-based tasks; DeepSeek-Coder-V2-Base, suitable for general text generation; and lightweight versions like DeepSeek-Coder-V2-Lite-Base and DeepSeek-Coder-V2-Lite-Instruct, designed for environments with limited computational resources. -
16
DeepInfer
DeepSeek
DeepInfer is an open-source inference framework for running DeepSeek language models on Linux systems with supported NVIDIA or Huawei Ascend accelerators. It provides offline inference and runtime components for model execution, optimized kernels, distributed parallelism, asynchronous batch scheduling, speculative decoding, deterministic generation, and structured outputs such as JSON Schema, regular expressions, and EBNF formats. The project documents DeepSeek-V4.1-Flash support and includes build instructions, configuration guidance, examples, and evaluation tools. It is intended for developers and infrastructure teams building or evaluating self-managed inference deployments. Its repository notes that production services may require additional KV-cache, prefill/decode disaggregation, and load-balancing components not included in this release. -
17
DeepSeek-V4.1-Flash
DeepSeek
DeepSeek-V4.1-Flash is a fast, versatile AI model designed for demanding coding, agentic, creative, and spatial reasoning workloads. Building on DeepSeek-V4-Flash, it emphasizes high-speed generation while maintaining strong performance on complex tasks, reaching more than 400 tokens per second and peaking at around 427 tokens per second in reported tests. The model can tackle advanced programming challenges, generate interactive 3D environments, create voxel-based designs, and reason about spatially complex scenes and simulations. Demonstrations include Minecraft-style worlds, classical Chinese gardens, racing environments, dungeon navigation, exploded camera views, and other applications requiring both code generation and an understanding of spatial relationships. Its capabilities make it suitable for rapid prototyping, game development, 3D workflows, architecture, research, and other technical or creative projects where iteration speed matters.Starting Price: Free -
18
DeepSeek
DeepSeek
DeepSeek is a cutting-edge AI assistant powered by the advanced DeepSeek-V3 model, featuring over 600 billion parameters for exceptional performance. Designed to compete with top global AI systems, it offers fast responses and a wide range of features to make everyday tasks easier and more efficient. Available across multiple platforms, including iOS, Android, and the web, DeepSeek ensures accessibility for users everywhere. The app supports multiple languages and has been continually updated to improve functionality, add new language options, and resolve issues. With its seamless performance and versatility, DeepSeek has garnered positive feedback from users worldwide.Starting Price: Free -
19
Qwen2.5-Max
Alibaba
Qwen2.5-Max is a large-scale Mixture-of-Experts (MoE) model developed by the Qwen team, pretrained on over 20 trillion tokens and further refined through Supervised Fine-Tuning (SFT) and Reinforcement Learning from Human Feedback (RLHF). In evaluations, it outperforms models like DeepSeek V3 in benchmarks such as Arena-Hard, LiveBench, LiveCodeBench, and GPQA-Diamond, while also demonstrating competitive results in other assessments, including MMLU-Pro. Qwen2.5-Max is accessible via API through Alibaba Cloud and can be explored interactively on Qwen Chat.Starting Price: Free -
20
DeepSeek-V3
DeepSeek
DeepSeek-V3 is a state-of-the-art AI model designed to deliver unparalleled performance in natural language understanding, advanced reasoning, and decision-making tasks. Leveraging next-generation neural architectures, it integrates extensive datasets and fine-tuned algorithms to tackle complex challenges across diverse domains such as research, development, business intelligence, and automation. With a focus on scalability and efficiency, DeepSeek-V3 provides developers and enterprises with cutting-edge tools to accelerate innovation and achieve transformative outcomes.Starting Price: Free -
21
Hy4
Tencent
Hy4 preview is a new-generation open source Mixture-of-Experts flagship model built for real-world productivity tasks across software engineering, office work, game development, and scientific research. The model contains 770B total parameters with 49B activated per token and supports a 1M-token context window, giving it the capacity to work through large codebases, extensive document collections, and long multi-step tasks. Its 78-layer architecture combines Gated DeepSeek Sparse Attention with IndexCache for cross-layer sparse index reuse and identity Hyper-Connections to expand information flow between layers. A native Multi-Token Prediction layer is included for speculative decoding. Hy4 preview is designed to understand, plan, debug, and verify long-horizon engineering tasks, with additional gains in front-end visual quality and interaction design. -
22
Kimi K2
Moonshot AI
Kimi K2 is a state-of-the-art open source large language model series built on a mixture-of-experts (MoE) architecture, featuring 1 trillion total parameters and 32 billion activated parameters for task-specific efficiency. Trained with the Muon optimizer on over 15.5 trillion tokens and stabilized by MuonClip’s attention-logit clamping, it delivers exceptional performance in frontier knowledge, reasoning, mathematics, coding, and general agentic workflows. Moonshot AI provides two variants, Kimi-K2-Base for research-level fine-tuning and Kimi-K2-Instruct pre-trained for immediate chat and tool-driven interactions, enabling both custom development and drop-in agentic capabilities. Benchmarks show it outperforms leading open source peers and rivals top proprietary models in coding tasks and complex task breakdowns, while its 128 K-token context length, tool-calling API compatibility, and support for industry-standard inference engines.Starting Price: Free -
23
GLM-5
Z.ai
GLM-5 is Z.ai’s latest large language model built for complex systems engineering and long-horizon agentic tasks. It scales significantly beyond GLM-4.5, increasing total parameters and training data while integrating DeepSeek Sparse Attention to reduce deployment costs without sacrificing long-context capacity. The model combines enhanced pre-training with a new asynchronous reinforcement learning infrastructure called slime, improving training efficiency and post-training refinement. GLM-5 achieves best-in-class performance among open-source models across reasoning, coding, and agent benchmarks, narrowing the gap with leading frontier models. It ranks highly on evaluations such as Vending Bench 2, demonstrating strong long-term planning and operational capabilities. The model is open-sourced under the MIT License.Starting Price: Free -
24
Phi-4-reasoning
Microsoft
Phi-4-reasoning is a 14-billion parameter transformer-based language model optimized for complex reasoning tasks, including math, coding, algorithmic problem solving, and planning. Trained via supervised fine-tuning of Phi-4 on carefully curated "teachable" prompts and reasoning demonstrations generated using o3-mini, it generates detailed reasoning chains that effectively leverage inference-time compute. Phi-4-reasoning incorporates outcome-based reinforcement learning to produce longer reasoning traces. It outperforms significantly larger open-weight models such as DeepSeek-R1-Distill-Llama-70B and approaches the performance levels of the full DeepSeek-R1 model across a wide range of reasoning tasks. Phi-4-reasoning is designed for environments with constrained computing or latency. Fine-tuned with synthetic data generated by DeepSeek-R1, it provides high-quality, step-by-step problem solving. -
25
Phi-4-reasoning-plus
Microsoft
Phi-4-reasoning-plus is a 14-billion parameter open-weight reasoning model that builds upon Phi-4-reasoning capabilities. It is further trained with reinforcement learning to utilize more inference-time compute, using 1.5x more tokens than Phi-4-reasoning, to deliver higher accuracy. Despite its significantly smaller size, Phi-4-reasoning-plus achieves better performance than OpenAI o1-mini and DeepSeek-R1 at most benchmarks, including mathematical reasoning and Ph.D. level science questions. It surpasses the full DeepSeek-R1 model (with 671 billion parameters) on the AIME 2025 test, the 2025 qualifier for the USA Math Olympiad. Phi-4-reasoning-plus is available on Azure AI Foundry and HuggingFace. -
26
Chat Stream
Chat Stream
Chat Stream provides access to two powerful language models from DeepSeek: Model Capabilities Utilizes DeepSeek V3 and R1 models with 671B parameters (37B activated per token) Achieves exceptional benchmark scores: MMLU (87.1%), BBH (87.5%) Features 128K context window length Supports code generation, mathematical computation, and multilingual processing Technical Features Advanced MoE (Mixture-of-Experts) architecture Multi-head Latent Attention (MLA) Auxiliary-loss-free load balancing Multi-token prediction objective Deployment Options Web-based chat interface with instant access One-click website integration via iframe Mobile apps for iOS and Android platforms Compatible with NVIDIA, AMD GPUs, and Huawei Ascend NPUs Supports both local inference and cloud deployment Access Methods Free chat access without registration Website embedding capabilities Mobile applications Premium subscription for ad-free experience -
27
Command A
Cohere AI
Command A, introduced by Cohere, is a high-performance AI model designed to maximize efficiency with minimal computational resources. This model outperforms or matches other top-tier models like GPT-4 and DeepSeek-V3 in agentic enterprise tasks while significantly reducing compute costs. It is tailored for applications requiring fast, efficient AI-driven solutions, providing businesses with the capability to perform advanced tasks across various domains, all while optimizing performance and computational demands.Starting Price: $2.50 / 1M tokens -
28
Beam
Reflection
Beam is Reflection’s first open-weight model, a sparse Mixture-of-Experts model with 501 billion total parameters and 23 billion active parameters, built for coding, reasoning, and agentic workloads. Its capabilities come from large-scale pretraining and reinforcement learning, including training on 23.8 trillion diverse, curated, high-quality tokens from the web, public sources, and proprietary licensed datasets. Beam was trained with a particular focus on coding and agentic performance and is designed to deliver competitive open-weight capabilities with efficient inference compute. It supports complex software engineering, terminal, STEM, web search, tool-use, and general knowledge tasks, with reinforcement learning designed to improve multi-step reasoning, tool use, and adaptation to environment feedback. Users can control the tradeoff between performance and token usage through a reasoning effort parameter. -
29
ModelArk
ByteDance
ModelArk is ByteDance’s one-stop large model service platform, providing access to cutting-edge AI models for video, image, and text generation. With powerful options like Seedance 1.0 for video, Seedream 3.0 for image creation, and DeepSeek-V3.1 for reasoning, it enables businesses and developers to build scalable, AI-driven applications. Each model is backed by enterprise-grade security, including end-to-end encryption, data isolation, and auditability, ensuring privacy and compliance. The platform’s token-based pricing keeps costs transparent, starting with 500,000 free inference tokens per LLM and 2 million tokens per vision model. Developers can quickly integrate APIs for inference, fine-tuning, evaluation, and plugins to extend model capabilities. Designed for scalability, ModelArk offers fast deployment, high GPU availability, and seamless enterprise integration. -
30
LongCat-2.0
LongCat
LongCat-2.0 is a 1.6 trillion total-parameter Mixture-of-Experts language model built on AI ASIC superpods, with about 48 billion parameters activated per token and strong performance across coding and agentic tasks. It is a substantial step up from previous LongCat models, combining large-scale sparse architecture with dedicated post-training for real-world software engineering, tool use, long-context reasoning, and multi-step agent workflows. LongCat-2.0 is trained and deployed entirely on AI ASIC superpods, with pretraining spanning more than 35 trillion tokens and millions of accelerator-hours, demonstrating frontier-scale training on alternative hardware platforms. To strengthen long-horizon tasks, the model introduces LongCat Sparse Attention and is trained on hundreds of billions of tokens of 1M-context data, giving it native support for ultra-long context tasks and reliable long-document understanding. -
31
Recursive Flint
Recursive
Recursive Flint is a 40-billion-parameter mixture-of-experts language model for agentic coding on local hardware. It activates about 2.5 billion parameters per token and supports reasoning, multi-turn tool use, and context windows up to 512,000 tokens. Developers can use Flint with coding agents to inspect codebases, fix bugs, implement features, and write tests while running inference on their own hardware. Its weights are published under the Apache 2.0 license. Flint works with llama.cpp, Ollama, Unsloth, LM Studio, vLLM, and SGLang; Recursive also provides rec, an inference engine optimized for Apple Silicon M4 Pro and M5 Pro/Max. Setup documentation and examples help developers run the model locally. Flint is for developers and technical teams that want to operate and adapt a code-focused model on their own systems. -
32
Vynaris
Vynaris
Vynaris gives teams powerful hosted uncensored models for authorized security testing, red-teaming, and research. Qwen3.8-27B, DeepSeek-V4-Flash-0731, and Qwen3.6-35B-A3B are available through an OpenAI-compatible API with published token rates and no prompt or output retention for these hosted models. Vynaris also routes requests across a broader model catalog and shows transparent per-request cost receipts, so applications can switch with a base URL change and inspect the cost of individual requests.Starting Price: $5 minimum credit top-up -
33
Qwen3
Alibaba
Qwen3, the latest iteration of the Qwen family of large language models, introduces groundbreaking features that enhance performance across coding, math, and general capabilities. With models like the Qwen3-235B-A22B and Qwen3-30B-A3B, Qwen3 achieves impressive results compared to top-tier models, thanks to its hybrid thinking modes that allow users to control the balance between deep reasoning and quick responses. The platform supports 119 languages and dialects, making it an ideal choice for global applications. Its pre-training process, which uses 36 trillion tokens, enables robust performance, and advanced reinforcement learning (RL) techniques continue to refine its capabilities. Available on platforms like Hugging Face and ModelScope, Qwen3 offers a powerful tool for developers and researchers working in diverse fields.Starting Price: Free -
34
Kimi K2 Thinking
Moonshot AI
Kimi K2 Thinking is an advanced open source reasoning model developed by Moonshot AI, designed specifically for long-horizon, multi-step workflows where the system interleaves chain-of-thought processes with tool invocation across hundreds of sequential tasks. The model uses a mixture-of-experts architecture with a total of 1 trillion parameters, yet only about 32 billion parameters are activated per inference pass, optimizing efficiency while maintaining vast capacity. It supports a context window of up to 256,000 tokens, enabling the handling of extremely long inputs and reasoning chains without losing coherence. Native INT4 quantization is built in, which reduces inference latency and memory usage without performance degradation. Kimi K2 Thinking is explicitly built for agentic workflows; it can autonomously call external tools, manage sequential logic steps (up to and typically between 200-300 tool calls in a single chain), and maintain consistent reasoning.Starting Price: Free -
35
Nemotron 3 Super
NVIDIA
Nemotron-3 Super is part of NVIDIA’s Nemotron 3 family of open models designed to enable advanced agentic AI systems that can reason, plan, and execute multi-step workflows across complex environments. The model introduces a hybrid Mamba-Transformer Mixture-of-Experts architecture that combines the efficiency of state-space Mamba layers with the contextual understanding of transformer attention, allowing it to process long sequences and complex reasoning tasks with high accuracy and throughput. This architecture activates only a subset of model parameters for each token, improving computational efficiency while maintaining strong reasoning capabilities and enabling scalable inference for large workloads. Nemotron-3 Super contains roughly 120 billion parameters with around 12 billion active during inference, accelerating multi-step reasoning and collaborative agent interactions across large contexts. -
36
Command A Translate
Cohere AI
Command A Translate is Cohere’s enterprise-grade machine translation model crafted to deliver secure, high-quality translation across 23 business-relevant languages. Built on a powerful 111-billion-parameter architecture with an 8K-input / 8K-output context window, it achieves industry-leading performance that surpasses models like GPT-5, DeepSeek-V3, DeepL Pro, and Google Translate across a broad suite of benchmarks. The model supports private deployments for sensitive workflows, allowing enterprises full control over their data, and introduces an innovative “Deep Translation” workflow, an agentic, multi-step refinement process that iteratively enhances translation quality for complex use cases. External validation from RWS Group confirms its excellence in challenging translation tasks. Additionally, the model’s weights are available for research via Hugging Face under a CC-BY-NC license, enabling deep customization, fine-tuning, and private deployment flexibility. -
37
Inkling-Small
Thinking Machines Lab
Inkling-Small is an efficient model that offers performance comparable to Inkling at a quarter of its size. It is a Mixture-of-Experts transformer with 276 billion total parameters and 12 billion active parameters, trained on NVIDIA GB300 NVL72 systems. It supports native reasoning across text, images, and audio, variable thinking effort, and context windows of up to one million tokens. Users adjust reasoning effort from minimal to extra high to balance performance and compute. Improved pre-training data, post-training with on-policy distillation from Inkling, and extended agentic coding reinforcement learning helped Inkling-Small surpass its larger counterpart on reasoning and coding benchmarks. It performs well in coding and tool-use harnesses, exceeds 80% on SWE-bench Verified, and combines strong reasoning with efficient output. Its encoder-free multimodal architecture processes audio as dMel spectrograms and images as 40-by-40-pixel patches alongside text tokens.Starting Price: $0.30 per million input tokens -
38
EXAONE Deep
LG
EXAONE Deep is a series of reasoning-enhanced language models developed by LG AI Research, featuring parameter sizes of 2.4 billion, 7.8 billion, and 32 billion. These models demonstrate superior capabilities in various reasoning tasks, including math and coding benchmarks. Notably, EXAONE Deep 2.4B outperforms other models of comparable size, EXAONE Deep 7.8B surpasses both open-weight models of similar scale and the proprietary reasoning model OpenAI o1-mini, and EXAONE Deep 32B shows competitive performance against leading open-weight models. The repository provides comprehensive documentation covering performance evaluations, quickstart guides for using EXAONE Deep models with Transformers, explanations of quantized EXAONE Deep weights in AWQ and GGUF formats, and instructions for running EXAONE Deep models locally using frameworks like llama.cpp and Ollama.Starting Price: Free -
39
GigaChat 3 Ultra
Sberbank
GigaChat 3 Ultra is a 702-billion-parameter Mixture-of-Experts model built from scratch to deliver frontier-level reasoning, multilingual capability, and deep Russian-language fluency. It activates just 36 billion parameters per token, enabling massive scale with practical inference speeds. The model was trained on a 14-trillion-token corpus combining natural, multilingual, and high-quality synthetic data to strengthen reasoning, math, coding, and linguistic performance. Unlike modified foreign checkpoints, GigaChat 3 Ultra is entirely original—giving developers full control, modern alignment, and a dataset free of inherited limitations. Its architecture leverages MoE, MTP, and MLA to match open-source ecosystems and integrate easily with popular inference and fine-tuning tools. With leading results on Russian benchmarks and competitive performance on global tasks, GigaChat 3 Ultra represents one of the largest and most capable open-source LLMs in the world.Starting Price: Free -
40
Qwen2
Alibaba
Qwen2 is the large language model series developed by Qwen team, Alibaba Cloud. Qwen2 is a series of large language models developed by the Qwen team at Alibaba Cloud. It includes both base language models and instruction-tuned models, ranging from 0.5 billion to 72 billion parameters, and features both dense models and a Mixture-of-Experts model. The Qwen2 series is designed to surpass most previous open-weight models, including its predecessor Qwen1.5, and to compete with proprietary models across a broad spectrum of benchmarks in language understanding, generation, multilingual capabilities, coding, mathematics, and reasoning.Starting Price: Free -
41
Qwen3.5
Alibaba
Qwen3.5 is a next-generation open-weight multimodal large language model designed to power native vision-language agents. The flagship release, Qwen3.5-397B-A17B, combines a hybrid linear attention architecture with sparse mixture-of-experts, activating only 17 billion parameters per forward pass out of 397 billion total to maximize efficiency. It delivers strong benchmark performance across reasoning, coding, multilingual understanding, visual reasoning, and agent-based tasks. The model expands language support from 119 to 201 languages and dialects while introducing a 1M-token context window in its hosted version, Qwen3.5-Plus. Built for multimodal tasks, it processes text, images, and video with advanced spatial reasoning and tool integration. Qwen3.5 also incorporates scalable reinforcement learning environments to improve general agent capabilities. Designed for developers and enterprises, it enables efficient, tool-augmented, multimodal AI workflows.Starting Price: Free -
42
ERNIE X1 Turbo
Baidu
ERNIE X1 Turbo, developed by Baidu, is an advanced deep reasoning AI model introduced at the Baidu Create 2025 conference. Designed to handle complex multi-step tasks such as problem-solving, literary creation, and code generation, this model outperforms competitors like DeepSeek R1 in terms of reasoning abilities. With a focus on multimodal capabilities, ERNIE X1 Turbo supports text, audio, and image processing, making it an incredibly versatile AI solution. Despite its cutting-edge technology, it is priced at just a fraction of the cost of other top-tier models, offering a high-value solution for businesses and developers.Starting Price: $0.14 per 1M tokens -
43
DeepSeekMath
DeepSeek
DeepSeekMath is a specialized 7B parameter language model developed by DeepSeek-AI, designed to push the boundaries of mathematical reasoning in open-source language models. It starts from the DeepSeek-Coder-v1.5 7B model and undergoes further pre-training with 120B math-related tokens sourced from Common Crawl, alongside natural language and code data. DeepSeekMath has demonstrated remarkable performance, achieving a 51.7% score on the competition-level MATH benchmark without external tools or voting techniques, closely competing with the likes of Gemini-Ultra and GPT-4. The model's capabilities are enhanced by a meticulous data selection pipeline and the introduction of Group Relative Policy Optimization (GRPO), which optimizes both mathematical reasoning and memory usage. DeepSeekMath is available in base, instruct, and RL versions, supporting both research and commercial use, and is aimed at those looking to explore or apply advanced mathematical problem-solving in AI contexts.Starting Price: Free -
44
DeepSeek Coder
DeepSeek
DeepSeek Coder is a cutting-edge software tool designed to revolutionize the landscape of data analysis and coding. By leveraging advanced machine learning algorithms and natural language processing capabilities, it empowers users to seamlessly integrate data querying, analysis, and visualization into their workflow. The intuitive interface of DeepSeek Coder enables both novice and experienced programmers to efficiently write, test, and optimize code. Its robust set of features includes real-time syntax checking, intelligent code completion, and comprehensive debugging tools, all designed to streamline the coding process. Additionally, DeepSeek Coder's ability to understand and interpret complex data sets ensures that users can derive meaningful insights and create sophisticated data-driven applications with ease.Starting Price: Free -
45
Nemotron 3.5 Lightning
NVIDIA
NVIDIA Nemotron 3.5 Lightning is an open 30B-parameter mixture-of-experts model with 3B active parameters, designed for high-volume, low-latency execution in long-running and always-on AI agents. Built for the execution layer of agentic systems, it handles frequent tasks such as tool calls, output validation, routine commands, and subagent delegation while larger reasoning models focus on planning and orchestration. Its MoE architecture activates only a fraction of parameters for each token, combining the capacity of a larger model with lower compute requirements. The model is trained for popular agent harnesses and supports speculative decoding through multi-token prediction, DFlash, and DSpark to improve inference speed across different serving scenarios. It is available with BF16 and NVFP4 checkpoints and can run from local systems such as DGX Spark and GeForce RTX hardware to data center environments. -
46
Qwen3.8-2.4T-A95B
Alibaba
Qwen3.8-2.4T-A95B is the largest open model in the Qwen3.8 family, bringing Qwen-Max-class capabilities to an open release. Built on the architectural foundation of Qwen3.5, it delivers substantial improvements across coding, professional work, research, and long-horizon agentic tasks, with a focus on carrying complex, multi-step work through to completion more reliably. The causal language model uses a mixture-of-experts architecture with 2.4 trillion total parameters and 95 billion activated parameters, including 512 experts with 10 routed and one shared expert active at a time. It supports a native context length of 262,144 tokens that can be extended to approximately 1.01 million tokens. Agent execution is strengthened through better autonomous planning and improved handling of environment feedback, while broader compatibility with popular agent harnesses and development tools simplifies integration into existing stacks. -
47
Llama 2
Meta
The next generation of our open source large language model. This release includes model weights and starting code for pretrained and fine-tuned Llama language models — ranging from 7B to 70B parameters. Llama 2 pretrained models are trained on 2 trillion tokens, and have double the context length than Llama 1. Its fine-tuned models have been trained on over 1 million human annotations. Llama 2 outperforms other open source language models on many external benchmarks, including reasoning, coding, proficiency, and knowledge tests. Llama 2 was pretrained on publicly available online data sources. The fine-tuned model, Llama-2-chat, leverages publicly available instruction datasets and over 1 million human annotations. We have a broad range of supporters around the world who believe in our open approach to today’s AI — companies that have given early feedback and are excited to build with Llama 2.Starting Price: Free -
48
DeepSeek-OCR
DeepSeek
DeepSeek-OCR is an open source model for Contexts Optical Compression, built to explore the boundaries of visual-text compression and investigate the role of vision encoders from an LLM-centric viewpoint. It is designed to compress long contexts through optical 2D mapping, using DeepEncoder as the core engine and DeepSeek3B-MoE-A570M as the decoder. DeepEncoder maintains low activations under high-resolution input while achieving high compression ratios, keeping the number of vision tokens manageable for document understanding. The model supports OCR and document parsing workflows for images and PDFs, with inference through vLLM or Transformers. Users can run image OCR with streaming output, process PDFs with high concurrency, or run batch evaluation for benchmarks. DeepSeek-OCR can convert documents to Markdown, perform free OCR without layouts, parse figures, describe images in detail, and locate referenced text inside an image.Starting Price: Free -
49
Janus-Pro-7B
DeepSeek
Janus-Pro-7B is an innovative open-source multimodal AI model from DeepSeek, designed to excel in both understanding and generating content across text, images, and videos. It leverages a unique autoregressive architecture with separate pathways for visual encoding, enabling high performance in tasks ranging from text-to-image generation to complex visual comprehension. This model outperforms competitors like DALL-E 3 and Stable Diffusion in various benchmarks, offering scalability with versions from 1 billion to 7 billion parameters. Licensed under the MIT License, Janus-Pro-7B is freely available for both academic and commercial use, providing a significant leap in AI capabilities while being accessible on major operating systems like Linux, MacOS, and Windows through Docker.Starting Price: Free -
50
Qwen3.8-Flash-Next
Alibaba
Qwen3.8-Flash-Next is an open-weight multimodal Mixture-of-Experts model and an early preview of the architecture planned for Qwen4. It systematically upgrades attention, residual connections, embeddings, and optimization to improve capability, computational efficiency, model capacity, and training stability. Its hybrid architecture combines Gated DeltaNet, which efficiently compresses historical information, with Qwen Sparse Attention, which selects important context at the micro-block level to reduce attention and indexing costs on long sequences. Gated Residual widens the residual stream into four branches and dynamically controls information flow across layers, while N-gram Embedding adds large-scale local-pattern memory with very little extra per-token computation and can be offloaded to host memory. The model uses a 125B-parameter main network plus 51B N-gram embedding parameters, while activating only 6B parameters per token.Starting Price: $2 per 1M (input)
