Structured outputs

Mistral Large 3 675B Instruct is MistralAI's Mistral Large model. It offers a 128K-token context window and scores 70.2 on τ-bench.

$0.500 in / $1.50 out / 1M tokens

6 routes · 1 batch · 3 cache

Mistral Small 3 is a 24B-parameter, latency-optimized model that Mistral released on January 30, 2025, as both pretrained and instruction-tuned open-weight checkpoints. Mistral aims it at fast conversational assistants, low-latency function calling, fine-tuning for domain specialists, and local inference; once quantized, it can run on a single RTX 4090 or a MacBook with 32GB of RAM.

$0.150 in / $0.600 out / 1M tokens

2 routes · 1 cache

Qwen3.5-9B is Alibaba's Qwen3.5 model with multimodal text and image input. It offers a 256K-token context window with weights openly available for self-hosting.

$0.100 in / $0.150 out / 1M tokens

5 routes

Together AI - Llama 3 8B Lite is Meta's Llama 3 model. It offers an 8K-token context window with weights openly available for self-hosting.

No tracked provider route

Efficient 4B parameter model from Google, available on Together AI. Gemma 3 nano-edge model optimized for low-latency inference.

No tracked provider route

Llama Guard 7B is a specialized content moderation model based on the Llama 2 architecture, designed to safeguard AI interactions. With 7 billion parameters, it excels in classifying and moderating both input prompts and output responses from large language models. The model employs a comprehensive risk taxonomy to identify various categories of harmful content, including violence, hate speech, and sexual content. Trained on diverse datasets, including prompts from the Anthropic dataset and in-house generated responses, Llama Guard 7B has demonstrated superior performance compared to industry-standard content moderation APIs. This makes it an invaluable tool for AI engineers focused on deploying safe and responsible AI systems. For more information, visit the model's page on Hugging Face .

$0.200 in / $0.200 out / 1M tokens

1 route

Llama Guard 3 8B is a specialized large language model developed by Meta for content safety classification. Fine-tuned from Llama 3.1, this 8-billion parameter model excels in moderation tasks, classifying both inputs and outputs across 14 hazard categories based on the MLCommons taxonomy. It supports multiple languages, including English, French, German, Hindi, Italian, Portuguese, Spanish, and Thai. Designed for AI engineers focusing on safe and responsible AI systems, Llama Guard 3 offers improved accuracy and reduced false positive rates in identifying unsafe content, making it a valuable tool for developing robust content moderation systems in conversational AI applications .

$0.484 in / $0.030 out / 1M tokens

4 routes

Xiaomi MiMo-V2.5 is the lower-cost native omnimodal sibling in the MiMo-V2.5 series. OpenRouter describes it as supporting text, image, audio, and video inputs with text output, Pro-level agentic performance at roughly half the inference cost, and improved multimodal perception over MiMo-V2-Omni. Xiaomi's official April 22 release page highlights MiMo-V2.5 alongside MiMo-V2.5-Pro in benchmark data and says the V2.5 series will be open-sourced soon; no public weights/license were verified at research time.

$0.140 in / $0.280 out / 1M tokens

2 routes · 2 cache

Xiaomi's April 22, 2026 public-beta flagship in the MiMo-V2.5 series. The official Xiaomi MiMo page describes MiMo-V2.5-Pro as its most capable model to date, focused on general agentic capability, complex software engineering, long-horizon tasks, and ultra-long-context instruction following. OpenRouter lists it as text-to-text with 1,048,576 token context, 131,072 max completion tokens, reasoning controls, tool use, and response_format support. Xiaomi says the V2.5 series will be open-sourced soon, but no public weights/license were verified at research time.

$0.435 in / $0.870 out / 1M tokens

3 routes · 3 cache

Xiaomi MiMo-V2.6-Flash is Xiaomi MiMo's September 21, 2026 efficiency-balanced open-weights MoE (309B total / 15B active), sibling of MiMo-V2.6-Pro. First-party HF checkpoint XiaomiMiMo/MiMo-V2.6-Flash-RL (MIT; ungated; 1M context) is the released Flash RL weights; product/API id is mimo-v2.6-flash. First-party news: the series comprises two native fully multimodal models (Pro and Flash); Flash comprehensively outperformed MiMo-V2.5-Pro. Product card: full-modality high-intelligence low-cost reasoning for high-frequency and large-scale professional workflows; context 1M; max output 128K; input text, image, video, and audio. Overseas API pricing (pay-as-you-go): $0.0028 cache-hit / $0.14 cache-miss input and $0.28 output per 1M tokens; batch $0.0014 / $0.07 / $0.14. Sparse MoE MiMoV2ForCausalLM; 48 layers (39 SWA / 9 GA); 256 routed experts / 8 activated per token; hybrid SWA+GA backbone; native omnimodal (text/image/video/audio); 681M MiMo ViT vision encoder; 5-layer SWA MTP/DFlash speculative decoder.

$0.140 in / $0.280 out / 1M tokens

2 routes · 1 batch · 2 cache

The Llama 2 7B Chat model is a fine-tuned variant of Meta's Llama 2 series, optimized for conversational AI applications. Built on an auto-regressive transformer architecture, it boasts 7 billion parameters and has been trained on a diverse dataset of 2 trillion tokens. The model underwent supervised fine-tuning and reinforcement learning with human feedback to enhance its performance in dialogue scenarios. It demonstrates competitive capabilities in terms of helpfulness and safety compared to both open-source and closed-source alternatives like ChatGPT and PaLM. Designed for commercial and research use, particularly in English language tasks, it's well-suited for developing chatbots, virtual assistants, and other interactive AI systems. More details can be found on its Hugging Face page .

$0.200 in / $0.200 out / 1M tokens

5 routes

The Llama 2 13B Chat model is a 13 billion parameter generative text model developed by Meta, optimized for conversational applications. Released on July 18, 2023, it's part of the Llama 2 family and excels in dialogue scenarios. The model leverages supervised fine-tuning (SFT) and reinforcement learning with human feedback (RLHF) to generate coherent and contextually relevant responses. Trained on 2 trillion tokens from diverse public sources, it outperforms many open-source chat models and matches popular closed-source models in helpfulness and safety. This model is ideal for AI engineers working on chatbots, virtual assistants, and customer service automation. For more details, visit the model's Hugging Face page [1].

$0.200 in / $0.200 out / 1M tokens

5 routes

Kimi K3 is Moonshot AI's 2.8-trillion-parameter flagship multimodal model for long-horizon coding, knowledge work, deep reasoning, and agentic workflows. It uses Kimi Delta Attention, Attention Residuals, and a sparse MoE design (16 of 896 experts active), supports a 1,048,576-token context window, text/image/video input, always-on reasoning, ToolCalls, strict JSON Schema structured output, automatic context caching, and partial mode through Moonshot's OpenAI-compatible API.

$3.00 in / $15.00 out / 1M tokens

3 routes · 2 cache

Holo3.1-35B-A3B is H Company's flagship open-weights 35B (3B active) sparse MoE computer-use VLM released June 1, 2026 under Apache 2.0. Achieves 79.3% on AndroidWorld and >25% improvement on the Holotab harness over Holo3. Supports native function-calling, FP8/NVFP4 quantization for DGX Spark, and local self-hosting.

$0.250 in / $1.80 out / 1M tokens

1 route

GLM-5.3-Flash is Z.ai's first native multimodal model in the GLM-5 series. It is a 320B-total / 18B-active Mixture-of-Experts model with a 1M-token context window and 131,072 max output tokens, supporting image, video, and file input plus function calling, tool use, structured outputs, reasoning, and prompt caching. MIT open weights are ungated on Hugging Face at zai-org/GLM-5.3-Flash. Official API model ID: glm-5.3-flash. The same model was previously tested anonymously as ox-alpha; that name is not a separate catalog slug.

$0.150 in / $0.500 out / 1M tokens

2 routes · 2 cache

Holo3.1-9B is H Company's 9B-parameter open-weights computer-use VLM released June 1, 2026 under Apache 2.0. Part of the Holo3.1 family. Achieves 71% on AndroidWorld. Supports native function-calling and structured JSON output for web, desktop, and mobile agent workflows.

No tracked provider route

CodeLlama 7B Python is a specialized variant of Meta's CodeLlama family, designed for Python programming tasks. With 7 billion parameters, it excels in code completion, infilling, and instruction following. The model utilizes an optimized auto-regressive transformer architecture and has been trained on diverse programming tasks. It's suitable for both commercial and research applications, offering AI engineers a powerful tool for enhancing productivity in Python-centric environments. For more details, visit the model's page on Hugging Face .

$0.200 in / $0.200 out / 1M tokens

3 routes

CodeLlama 7B is a specialized code generation model released by Meta on August 24, 2023. With 7 billion parameters, it excels in code completion, infilling, and instruction-following tasks. Built on an optimized transformer architecture, it's designed for both commercial and research applications, particularly in English and various programming languages. The model offers robust capabilities for AI engineers looking to enhance coding workflows or develop code generation applications. More details can be found on the official Hugging Face page .

$0.200 in / $0.200 out / 1M tokens

5 routes

GLM-5.3 is Z.ai's August 2026 coding-first post-training release on the same GLM-5.2 base model, focused on complex software engineering, long-horizon agentic coding, and emergent cybersecurity evaluation. It is available through the GLM Coding Plan and Z.ai API; open weights are scheduled after safety hardening. Official API model ID: glm-5.3.

$1.40 in / $4.40 out / 1M tokens

3 routes · 2 cache

Holo3.1-4B is H Company's 4B-parameter open-weights computer-use VLM released June 1, 2026 under Apache 2.0. Part of the Holo3.1 family. Achieves 71% on AndroidWorld (up from 58% for Holo3). Supports native function-calling, web/desktop/mobile environments, and local deployment via quantized checkpoints.

No tracked provider route

CodeLlama 13B Python is a specialized variant of the CodeLlama family, featuring 13 billion parameters and optimized for Python programming tasks. This model excels at code synthesis, completion, and infilling, while also supporting instruction following and chat-based interactions. Leveraging advanced transformer architecture and trained on a diverse dataset, it offers robust understanding of programming concepts and syntax. Designed for AI engineers and developers, it serves as a powerful tool for integrating AI into coding workflows, making it particularly valuable for Python-related applications .

$0.200 in / $0.200 out / 1M tokens

3 routes

CodeLlama 13B is a state-of-the-art generative text model developed by Meta, specifically designed for code synthesis and understanding tasks. Released on August 24, 2023, this 13-billion-parameter model excels in general code generation and comprehension, making it suitable for a wide range of programming tasks, including code completion, infilling, and instruction following. It utilizes an optimized transformer architecture and has been trained on a diverse dataset similar to Llama 2, ensuring robust understanding of programming languages and coding practices. AI engineers can integrate CodeLlama 13B into various coding environments and tools for both commercial and research applications, leveraging its powerful capabilities to enhance productivity and streamline the coding process.

$0.200 in / $0.200 out / 1M tokens

4 routes

GLM-5.2 is Z.ai's coding-first successor to GLM-5.1 in the GLM-5 family, released June 13 2026. 753B parameters (40B active) in IndexShare MoE architecture; the IndexShare innovation reuses the same attention indexer across every four sparse layers, cutting per-token FLOPs by 2.9x at 1M context length. Trained on 28.5T tokens. Supports a 1M-token context window via the glm-5.2[1m] model ID, with 131,072-token maximum output and High/Max thinking-effort levels designed for extended agentic coding sessions. MIT license; open weights available on Hugging Face (zai-org/GLM-5.2 and zai-org/GLM-5.2-FP8). Self-reported HF card benchmarks: SWE-bench Pro 62.1, Terminal-Bench 2.1 82.7, MCP-Atlas 76.8, Tool-Decathlon 48.2, GPQA Diamond 91.2, AIME 2026 99.2, HLE 40.5. Available to GLM Coding Plan subscribers (Lite/Pro/Max/Team) directly, and via OpenRouter token API ($1.40/$4.40 per 1M tokens).

$1.40 in / $4.40 out / 1M tokens

2 routes · 1 cache

Holo3.1-0.8B is H Company's smallest 0.8B-parameter open-weights computer-use VLM, released June 1, 2026 under Apache 2.0. Part of the Holo3.1 family spanning 0.8B-35B. Fine-tuned from Qwen3.5-0.8B. Supports native function-calling and structured JSON output for local/edge deployment.

No tracked provider route

CodeLlama 34B Python is a specialized code generation model released by Meta on August 24, 2023. With 34 billion parameters, it excels in Python-specific tasks like code completion, infilling, and instruction following. This model offers AI engineers a powerful tool for enhancing coding workflows and productivity. Its architecture is optimized for understanding and generating complex code structures, making it particularly useful for software development tasks. The model is available in the Hugging Face Transformers format, facilitating easy integration into existing projects .

$0.900 in / $0.900 out / 1M tokens

3 routes

Post-training variant of GLM-5 from Z.ai (Zhipu AI) with enhanced agentic coding capabilities. Released April 7, 2026. 754B parameters (40B active) in Mixture of Experts architecture, 200K token context, 128K max output. Supports autonomous plan–execute–test–fix–optimize loops for up to 8 hours without human intervention. Trained entirely on Huawei Ascend hardware (no Nvidia). Key benchmarks: SWE-bench Pro 58.4 (world #1 at release, surpassing GPT-5.4 57.7 and Claude Opus 4.6 57.3), GPQA Diamond 86.2, AIME 2026 95.3, Terminal-Bench 2.0 63.5, MCP-Atlas 71.8, Chatbot Arena Elo 1475 (June 16, 2026, arena.ai). Available via Z.ai API ($1.40/$4.40 per 1M input/output tokens) and open weights on Hugging Face under MIT license.

$1.38 in / $4.40 out / 1M tokens

5 routes · 3 cache

CodeLlama 34B is a powerful generative text model developed by Meta, specifically tailored for code synthesis and understanding. With 34 billion parameters, it excels in code completion, infilling, and instruction following, particularly for Python programming. The model utilizes an auto-regressive transformer architecture and has been trained on a diverse dataset of programming languages, making it versatile for various coding tasks. Designed for both commercial and research applications, CodeLlama 34B offers AI engineers a robust tool for integrating advanced code generation capabilities into their projects. More details can be found on the model's Hugging Face page .

$0.800 in / $0.800 out / 1M tokens

5 routes

Holo3-122B-A10B is H Company's flagship 122B (10B active) sparse MoE computer-use VLM. Released March 31, 2026, available via the H Company API only (no public open weights). Achieves 78.85% on OSWorld-Verified - SOTA at launch. Paid-tier API access at $0.40/$3.00 per 1M input/output tokens.

$0.400 in / $3.00 out / 1M tokens

1 route

Grok-2 is xAI's Grok 2 model with an optional reasoning mode. It offers a 128K-token context window.

$0.500 in / $0.500 out / 1M tokens

1 route

Qwen1.5-0.5B is Alibaba's Qwen1.5 model. It was released 2024-02-05.

$0.050 in / $0.250 out / 1M tokens

1 route

Qwen1.5-1.8B is Alibaba's Qwen1.5 model. It was released 2024-02-05.

$0.050 in / $0.250 out / 1M tokens

2 routes

Qwen1.5-4B is Alibaba's Qwen1.5 model. It was released 2024-02-05.

$0.050 in / $0.250 out / 1M tokens

1 route

Qwen1.5-7B is Alibaba's Qwen1.5 model. It scores 42.5 on the GPQA benchmark.

$0.050 in / $0.250 out / 1M tokens

2 routes

Gemma 7B is Google DeepMind's Gemma model. It offers an 8K-token context window with weights openly available for self-hosting.

$0.200 in / $0.200 out / 1M tokens

2 routes

Mistral Small is a language model from MistralAI. It offers a 32K-token context window.

$1.00 in / $3.00 out / 1M tokens

2 routes

Qwen1.5-32B is a robust large language model from the Qwen1.5 series, serving as a beta version of Qwen2. It is a transformer-based, decoder-only model, pretrained on an extensive dataset. Key features include its 32 billion parameters and a support for up to 32K context length, alongside multilingual capabilities. The model demonstrates substantial performance enhancements over its predecessor, especially in chat applications, using advanced techniques like SwiGLU activation and group query attention. While there's a base version, the chat variant is fine-tuned for conversational AI. It's accessible on Hugging Face and other platforms for diverse applications.

$0.200 in / $1.00 out / 1M tokens

1 route

Gemma 7B Instruct is a cutting-edge large language model developed by Google DeepMind, boasting 7 billion parameters. As part of the Gemma family, it benefits from the advanced research underpinning Google's Gemini models. This model is optimized for text generation tasks, excelling in areas like question answering and summarization, and it is finely tuned to follow instructions effectively. Despite its compact size, Gemma 7B Instruct performs impressively on benchmarks, making it versatile for deployment across various hardware platforms, from laptops to cloud infrastructure. Moreover, it is open-source, with accessible weights and incorporates responsible AI practices, such as data filtering and human feedback, to ensure safe and ethical use.

$0.200 in / $0.200 out / 1M tokens

4 routes

Qwen1.5-72B is Alibaba's Qwen1.5 model. It scores 79.8 on the Massive Multitask Language Understanding benchmark.

$0.900 in / $0.900 out / 1M tokens

3 routes

CodeLlama 70B Python is a specialized AI model by Meta, designed for Python code synthesis and understanding. With 70 billion parameters, it excels in code completion, infilling, and instruction following tasks. The model leverages an optimized transformer architecture and has been fine-tuned with up to 16,000 tokens, making it particularly effective for Python-centric development workflows. While it doesn't support long contexts of 100,000 tokens, it offers powerful capabilities for both commercial and research applications in Python programming environments. More details can be found in the research paper "Code Llama: Open Foundation Models for Code" .

$0.900 in / $0.900 out / 1M tokens

3 routes

Falcon-7B, developed by the Technology Innovation Institute, is a cutting-edge large language model boasting a decoder-only architecture with 7 billion parameters. It's trained on 1,500 billion tokens from the curated web dataset, RefinedWeb, enhancing its performance in language tasks. The model is equipped with advanced features like FlashAttention and multiquery attention, optimizing speed and memory usage. With 32 layers and rotary positional embeddings, it manages a sequence length of up to 2048 tokens efficiently. Renowned for tasks such as text generation, summarization, translation, and conversational AI, Falcon-7B is open-source under Apache 2.0, suitable even for consumer hardware, needing at least 16GB of memory for inference 236.

$0.520 in / $0.670 out / 1M tokens

3 routes

Nova Pro is Amazon's Nova model. It offers a 300K-token context window.

$0.800 in / $3.20 out / 1M tokens

2 routes

Reka Edge is Reka's Reka model with multimodal text and image input. It offers a 64K-token context window.

$0.100 in / $0.100 out / 1M tokens

2 routes

Mistral Large 2.1 is Mistral AI's November 2024 flagship Large 2 snapshot for high-complexity reasoning, coding, and function-calling workloads, with a 128K context window.

Pricing not tracked / 1M tokens

1 route

The Nous Hermes Llama 2 7B is a state-of-the-art large language model built on the efficient Llama 2 transformer architecture. Fine-tuned on over 300,000 instructions, it exhibits several impressive features, such as generating long, detailed responses with a low hallucination rate. Notably, it lacks OpenAI's censorship, enabling more open discussions. The model excels in knowledge retention and task completion through extensive training on synthetic GPT-4 outputs and supports prompts in the versatile Alpaca format. Its benchmark performance varies across tasks like GPT4All and BigBench, and quantized versions are available, providing flexible deployment across various platforms.

$0.200 in / $0.200 out / 1M tokens

1 route

Amazon Titan Text Lite is a lightweight and efficient large language model designed specifically for English-language tasks. It excels in fine-tuning applications such as summarization and copywriting, providing a cost-effective and highly customizable solution for users. Despite being smaller and less expensive than other Titan Text models, it supports a variety of text generation tasks. The model can handle a maximum context length of 4,000 tokens, ensuring flexibility in handling longer text inputs while maintaining efficiency and performance.

$0.150 in / $0.200 out / 1M tokens

1 route

The WizardLM 13B V1.0 is a transformer-based large language model built on the LLaMA architecture, featuring 13 billion parameters, enabling it to efficiently process a variety of natural language tasks. Its capabilities include text generation, coherent conversations, summarization, translation, and coding assistance. Trained on diverse datasets, the model has some limitations, such as potential biases, lack of content filtering, performance trade-offs due to quantization, and a context length limited to 2048 tokens. Despite these constraints, it is a powerful tool for numerous applications in natural language processing.

$0.300 in / $0.300 out / 1M tokens

1 route

WizardLM-2 7B is a large language model developed by WizardLM in collaboration with Microsoft AI. It is part of the WizardLM-2 family, which includes larger models but is notable for its quick processing speed, achieving performance comparable to open-source models that are much larger. This multilingual model can process diverse input types, such as natural language text, code, and mathematical expressions. It showcases capabilities in text generation, question answering, summarization, as well as code generation and mathematical problem-solving. Particularly adept at complex chat scenarios and multilingual tasks, it is based on the Mistral-7B-v0.1 base model and is available as open-source under the Apache 2.0 license. Different versions exist with various quantizations, providing options that balance model size and performance.

No tracked provider route

The Llama 3 8B Instruct model, released on April 18, 2024, is Meta's latest instruction-following language model with 8 billion parameters. It utilizes an auto-regressive transformer architecture with Grouped-Query Attention for improved scalability. Trained on over 15 trillion tokens and fine-tuned with 10 million human-annotated examples, it excels in dialogue and conversational tasks. The model outperforms its predecessors on industry benchmarks, scoring 68.4 on MMLU (5-shot). Designed for commercial and research applications, it prioritizes safety and helpfulness, making it suitable for chatbots, virtual assistants, and other interactive AI applications. For more details, visit the Hugging Face page [1].

$0.200 in / $0.200 out / 1M tokens

7 routes

Falcon 40B is a leading open-source large language model developed by the Technology Innovation Institute in Abu Dhabi, featuring a causal decoder-only architecture with 40 billion parameters. It stands out with its use of rotary positional embeddings, multi-query attention, and FlashAttention, enhancing its contextual understanding and processing efficiency. Trained on 1 trillion tokens using the enriched RefinedWeb dataset, Falcon 40B excels in various natural language processing tasks, ranging from text generation to language translation and question answering. It supports multiple languages and is open under the Apache 2.0 license, promoting both research and commercial use. The model efficiently utilizes standard hardware, requiring around 85-100 GB of memory for inference, setting a benchmark for performance and scalability in its category.

$1.54 in / $1.77 out / 1M tokens

4 routes

Mistral Large 2 (version 24.07, API name mistral-large-2407) is a 123-billion-parameter dense model that Mistral released on July 24, 2024, designed for single-node inference in long-context applications. It has a 128K-token context window, up from 32K on the first Mistral Large, and supports dozens of languages, including French, German, Spanish, Italian, Portuguese, Arabic, Hindi, Russian, Chinese, Japanese and Korean, plus more than 80 coding languages such as Python, Java, C, C++, JavaScript and Bash. Mistral trained it on a large proportion of code and reports stronger code and reasoning than the first Mistral Large, and the model supports native function calling, including parallel and sequential calls, and JSON output. The instruct weights are on Hugging Face under the Mistral Research License, which allows research and non-commercial use; commercial self-deployment requires a Mistral Commercial License. Mistral retired mistral-large-2407 from its API on March 30, 2025, and lists Mistral Large 3 as the replacement.

$2.00 in / $6.00 out / 1M tokens

3 routes · 1 batch · 1 cache

Nova Lite is Amazon's Nova model. It offers a 300K-token context window.

$0.060 in / $0.240 out / 1M tokens

2 routes

Jamba 1.5 Mini is AI21 Labs' Jamba 1.5 model. It offers a 256K-token context window.

$0.200 in / $0.400 out / 1M tokens

2 routes

The Nous Hermes Llama 2 13B is an advanced large language model fine-tuned on over 300,000 instructions, building upon the Llama 2 architecture. Developed by Nous Research with contributions from Teknium, Emozilla, and compute sponsorship from Redmond AI, the model boasts features like extended response generation, reduced hallucinations, and compatibility with Alpaca prompt format. It is trained on a diverse dataset, including synthetic GPT-4 outputs, enhancing its knowledge and task completion capabilities. Known for strong performance across various benchmarks, it is suitable for applications like chatbots, content generation, and code generation, while still acknowledging limitations such as data bias and handling of open-ended prompts.

$0.200 in / $0.200 out / 1M tokens

2 routes

Jurassic-2 Mid, developed by AI21 Labs, is a large language model that balances quality, speed, and cost, making it well-suited for complex language tasks such as chatbots and conversational interfaces. It has a parameter size of 17 billion and supports multiple languages including Spanish, French, German, Portuguese, Italian, and Dutch. Optimized for generating precise text from instruction-only prompts, the model is capable of zero-shot text generation without requiring examples. Despite its powerful capabilities, it shares common limitations with other LLMs, such as potential inaccuracies, lack of coherence, and the presence of training data biases. It was trained on a vast corpus of roughly 1.2 trillion tokens from diverse sources like CommonCrawl, Wikipedia, and Stack Exchange. Jurassic-2 Mid has an end-of-life date set for November 14, 2024 12.

$12.50 in / $12.50 out / 1M tokens

2 routes

The Qwen1.5-110B is a large language model created by Alibaba Cloud, distinguished as the largest in the Qwen1.5 series. It is a transformer-based, decoder-only model equipped with 110 billion parameters and optimized for efficiency using features like SwiGLU activation and Grouped Query Attention (GQA). Pretrained on an extensive dataset, it supports a 32K context length and multilingual capabilities, handling various languages including English and Chinese. The model excels in tasks like text generation, dialogue systems, and is noted for its competitive performance and advanced tokenizer, making it highly versatile and applicable across multiple NLP tasks. Various quantized versions are available to accommodate different hardware specifications.

$1.50 in / $2.50 out / 1M tokens

1 route

OpenChat 3.5 is an open-source large language model renowned for delivering performance on par with ChatGPT despite having just 7 billion parameters compared to ChatGPT's 70 billion. Built on a transformer architecture, this model uses C-RLFT (Constrained Reinforcement Learning from Trajectories) to learn from mixed-quality data without requiring explicit preference labels. It adeptly handles tasks such as text and code generation, including languages like HTML5 and JavaScript, and excels in multi-turn conversations. OpenChat 3.5 is open-source, making it an appealing choice for developers and researchers, and it is optimized to run efficiently on consumer-grade GPUs with 24GB of RAM. The model also includes specialized modes like "Mathematical Reasoning Mode" for task-specific enhancements. However, it is not without its challenges, as it shares common LLM limitations, such as hallucination and potential safety concerns related to biased or harmful outputs.

Pricing not tracked / 1M tokens

1 route

Amazon Titan Text Express is a large language model (LLM) crafted by AWS, offering a balance of price and performance for text generation 123. As part of the Amazon Titan family, this versatile model excels in tasks such as open-ended text generation, conversational chat, and Retrieval Augmented Generation (RAG) 124. It supports a context length of up to 8,000 tokens, enabling it to handle extensive text inputs effectively 235. While optimized for English, it provides preview support for over 100 other languages 235. This model can be fine-tuned with user data to enhance accuracy for specific tasks like summarization, code generation, table creation, data formatting, paraphrasing, and question answering 123.

$0.200 in / $0.600 out / 1M tokens

1 route

The Phi-3 Medium 4K, developed by Microsoft, is a state-of-the-art large language model with 14 billion parameters. It is engineered for efficiency across various tasks, particularly excelling in reasoning capabilities. This model is designed to handle 4,096 token context lengths, allowing for the processing of longer input sequences. Leveraging a dense, decoder-only Transformer architecture, it incorporates techniques like supervised fine-tuning and direct preference optimization to align with human preferences and safety standards. The model supports multilingual data, although it is primarily trained in English. Its lightweight nature allows for deployment on diverse hardware platforms, making it accessible and versatile for both commercial and research purposes. Safety measures are embedded, although further precautions are advised for applications with higher risks.

No tracked provider route

DeepSeek LLM 67B Chat is a sophisticated language model with 67 billion parameters, leveraging the LLaMA architecture with enhancements such as Grouped-Query Attention across 95 layers. Trained on a vast corpus of 2 trillion tokens in English and Chinese, it excels in tasks like text generation, question answering, and fluent conversation, demonstrating superior performance in reasoning, coding, and mathematics compared to some larger models. Despite its advanced capabilities, the model can exhibit biases from its training data, experience hallucinations, and produce repetitive outputs. Due to its size, substantial computational resources are needed for inference, although quantization methods can reduce its size with potential trade-offs in quality.

No tracked provider route

The Llama 3 70B Instruct model is a large language model with 70 billion parameters, released by Meta on April 18, 2024. It's an instruction-tuned variant optimized for conversational applications, utilizing an advanced auto-regressive transformer architecture. The model excels in following instructions and engaging in dialogue, having been trained on over 15 trillion tokens with a December 2023 knowledge cutoff. It demonstrates superior performance on industry benchmarks, scoring 82.0 on the MMLU (5-shot) test. The model incorporates extensive safety measures and optimizations, including RLHF, to enhance helpfulness and reduce harmful content generation. For more details, visit the model's Hugging Face page [1].

$0.400 in / $0.400 out / 1M tokens

10 routes

DeepSeek R1: Reasoning-optimized model with extended thinking capabilities. 128K context.

$0.100 in / $0.300 out / 1M tokens

9 routes

DeepSeek V3: Latest flagship model. 685B total with MoE. 128K context. Open-source.

$0.100 in / $0.300 out / 1M tokens

9 routes

Kimi K2.5 is Moonshot AI's Kimi model focused on code generation and software engineering. It offers a 256K-token context window and scores 87.9 on GPQA.

$0.440 in / $2.00 out / 1M tokens

7 routes · 1 cache

Flagship open-weight foundation model from Zhipu AI with 744B parameters (40B active per token) in Mixture of Experts architecture. Trained on 28.5T tokens using DeepSeek Sparse Attention on Huawei Ascend hardware. Achieves state-of-the-art performance on coding and agentic benchmarks (SWE-bench Verified: 77.8%). Supports autonomous planning, multi-step tool use, and self-correction.

$1.00 in / $3.20 out / 1M tokens

7 routes · 3 cache

Instruction-optimized flagship variant for demanding production applications requiring high-accuracy complex problem-solving across industries.

$0.280 in / $0.280 out / 1M tokens

7 routes

Kimi K3.1 Pro is the first model in Moonshot AI's K3.1 series, a post-trained version of Kimi K3. It has a 1,048,576-token context window, always-on reasoning with low, high and max effort, image and video input, tool calling and strict JSON Schema structured output. Released 2026-10-23 on Kimi.ai, Kimi Work, Kimi Code and the Kimi API; Moonshot published the weights on Hugging Face under the Kimi K3.1 Pro License.

$2.00 in / $10.00 out / 1M tokens

6 routes · 5 cache

DeepSeek V4 Pro is DeepSeek's flagship open-weights model, released April 24 2026 under the MIT license. Architecture: 1.6T total / 49B active parameters, MoE with Compressed Sparse Attention (CSA) + Heavily Compressed Attention (HCA) hybrid — requiring only 27% of inference FLOPs vs standard 1M-context transformers — plus Manifold-Constrained Hyper-Connections (mHC) and Muon Optimizer. Context window: 1,000,000 tokens; max output: 384,000 tokens (Think Max mode requires >=384K context). Text-only (no vision/image input). Supports three reasoning modes: Non-Think, Think High, Think Max. Function calling, tool use, and structured outputs supported. Key benchmarks: SWE-bench Verified 80.6%, SWE-bench Pro 55.4%, LiveCodeBench 93.5%, GPQA Diamond 90.1%, MMLU-Pro 87.5% and Chatbot Arena 1456 (2026-06-16).

$0.3633 in / $0.7266 out / 1M tokens

6 routes · 4 cache

NVIDIA Nemotron 3 Super-120B-A12B is a 120B total / 12B active hybrid Latent MoE model with interleaved Mamba-2 and MoE layers for agentic, reasoning, and conversational tasks. Fireworks lists the NVFP4 variant for on-demand deployment with 262k context.

$0.085 in / $0.400 out / 1M tokens

6 routes · 2 cache

Kimi K2 Instruct is an instruction-tuned language model from Moonshot AI, available via Fireworks AI.

$0.570 in / $2.30 out / 1M tokens

6 routes

Instruction-optimized 32B code flagship for production systems requiring top-tier code reasoning, generation, and multi-file analysis.

$0.180 in / $0.180 out / 1M tokens

6 routes

Llama 3.2 1B Instruct is Meta's Llama 3.2 model. It offers a 128K-token context window with weights openly available for self-hosting and scores 25.6 on GPQA.

$0.100 in / $0.100 out / 1M tokens

6 routes

Llama 3.2 3B Instruct is Meta's Llama 3.2 model. It offers a 128K-token context window with weights openly available for self-hosting and scores 34.7 on MMLU PRO.

$0.030 in / $0.050 out / 1M tokens

6 routes

Gemma 2 27B Instruct is a cutting-edge large language model from Google, excelling in text generation, question answering, summarization, and reasoning tasks. It features a decoder-only transformer architecture, utilizing 27 billion parameters, and supports context length processing of up to 8,192 tokens. The model incorporates innovative mechanisms like Grouped Query Attention and Sliding Window Attention to enhance efficiency and effectiveness in handling long texts. Its instruction-tuned variants are designed for improved interaction in conversational tasks, and it benefits from knowledge distillation techniques for enhanced performance. Additionally, Gemma 2 27B Instruct is openly accessible, promoting wider innovation in AI applications.

$0.400 in / $0.400 out / 1M tokens

6 routes

Instruction-tuned 7B variant combining strong reasoning with real-time inference on single GPUs, ideal for developer tools and vision applications.

$0.050 in / $0.050 out / 1M tokens

6 routes

Alibaba's closed-weight flagship language model, announced at the 2026 Alibaba Cloud Summit (May 20). 1M-token context with prompt caching.

$1.25 in / $3.75 out / 1M tokens

4 routes · 2 cache

Phi-4 14B is Microsoft Research's Phi-4 model. Weights are openly available for self-hosting and scores 56.1 on GPQA.

$0.070 in / $0.140 out / 1M tokens

4 routes

Instruction-tuned 11B Llama 3.2 Vision model for image reasoning, visual question answering, document understanding, and captioning. NVIDIA NIM lists text plus image input, text output, and a 128K context window for the Llama 3.2 Vision collection.

$0.200 in / $0.200 out / 1M tokens

4 routes

Instruction-tuned 32B variant for advanced content creation, analytics, and vision-language pipelines on multi-GPU infrastructure.

$0.150 in / $0.150 out / 1M tokens

4 routes

joey-1 is a 2B open-weight decision model from Base Labs, Baseten's research lab, released October 29, 2026. It is a fine-tune of Qwen3.5-2B that turns a state and typed questions into calibrated probabilities for choice, yes/no and score answers in a single forward pass, without generating free-form text, for routing, classification and rubric scoring. It accepts text and images and has a 16,384-token context limit. It is the smallest of Base Labs' three decision models (with elephant-1 and mammoth-1). Weights are on Hugging Face at baselabs/joey-1 under MIT.

$0.010 in / Free out / 1M tokens

3 routes · 1 cache

GLM-Legal is Z.ai's legal-domain model, built on GLM-5.3-Flash with legal mid-training and reinforcement-learning post-training across legal agent environments. Z.ai reports a 19.2% all-pass rate on the Harvey Legal Agent Benchmark, 12.5 points above its base model, as evaluated by Vals.

$0.370 in / $1.25 out / 1M tokens

3 routes · 3 cache