DeepSeek-R1 (DeepSeek)
DeepSeek-R1 introduces autonomous reasoning behaviors through pure reinforcement learning (DeepSeek-R1-Zero) followed by multi-stage refinement. It achieves competitive mathematical, coding, and logical reasoning benchmarks comparable to proprietary closed-source reasoning systems while providing full model weights under an MIT license.
Parameters: 671B MoE (37B active per token) | Context: 128,000 tokens | License: MIT
Official DocumentationOpenAI o1 (OpenAI)
OpenAI o1 utilizes reinforcement learning during pre-training and test-time compute to generate private internal chains of thought. It specializes in competitive programming, physics, PhD-level biology, and multi-step theorem verification.
Parameters: Undisclosed | Context: 200,000 tokens (up to 100k completion) | License: Proprietary Commercial API
Official DocumentationOpenAI o3-mini (OpenAI)
Engineered specifically for STEM disciplines and technical coding workflows, o3-mini offers controllable reasoning effort (low, medium, high) allowing developers to balance latency, cost, and analytical depth for production software environments.
Parameters: Undisclosed (Compact) | Context: 200,000 tokens | License: Proprietary Commercial API
Official DocumentationGemini 2.0 Flash (Google DeepMind)
Gemini 2.0 Flash is natively multimodal from inception across text, high-resolution audio, vision, and video. It delivers sub-second time-to-first-token latency, native bidirectional audio streaming, and seamless integration with external search grounding and code execution.
Parameters: Undisclosed | Context: 1,048,576 tokens | License: Proprietary Commercial API / Google AI Studio
Official DocumentationGemini 2.0 Pro Experimental (Google DeepMind)
Gemini 2.0 Pro Experimental achieves state-of-the-art results across code benchmarks and complex data analysis, offering 2 million tokens of context window capable of ingesting massive codebases and multi-hour video archives with high needle-in-a-haystack recall.
Parameters: Undisclosed | Context: 2,097,152 tokens | License: Proprietary Commercial API / Google AI Studio
Official DocumentationGemini 1.5 Pro (Google DeepMind)
Gemini 1.5 Pro broke context window barriers with its 2-million-token capacity and Sparse Mixture-of-Experts architecture. It processes an hour of video, eleven hours of audio, or over 700,000 words in a single query with near-perfect retrieval.
Parameters: Undisclosed Sparse MoE | Context: 2,000,000 tokens | License: Proprietary Commercial API
Official DocumentationClaude 3.5 Sonnet (Anthropic)
Claude 3.5 Sonnet operates at twice the speed of Claude 3 Opus while outperforming top models on graduate-level reasoning, undergraduate-level knowledge, and software coding benchmarks. It pioneered the Computer Use API for direct OS GUI interaction.
Parameters: Undisclosed | Context: 200,000 tokens | License: Proprietary Commercial API
Official DocumentationClaude 3.5 Haiku (Anthropic)
Claude 3.5 Haiku is Anthropic's fastest model. Designed for high-volume enterprise tasks, user-facing chat, and agentic sub-tasks, it matches or surpasses the original Claude 3 Opus on multiple engineering benchmarks at fraction of the latency.
Parameters: Undisclosed (Compact) | Context: 200,000 tokens | License: Proprietary Commercial API
Official DocumentationGPT-4o (Omni) (OpenAI)
GPT-4o trains a single neural network end-to-end across text, vision, and audio, preserving emotional inflections, background acoustics, and conversational interruptions that were lost in cascaded voice pipelines.
Parameters: Undisclosed | Context: 128,000 tokens | License: Proprietary Commercial API
Official DocumentationLlama 3.3 70B Instruct (Meta AI)
Meta's Llama 3.3 70B incorporates post-training advancements, knowledge distillation, and online DPO to deliver industry-leading performance in a dense 70B footprint capable of running on a single modern GPU server node.
Parameters: 70.6 Billion Dense | Context: 128,000 tokens | License: Llama 3.3 Community License
Official DocumentationLlama 3.1 405B (Meta AI)
Trained on over 16,000 H100 GPUs, Llama 3.1 405B represents a milestone for open science, providing a premier open baseline for synthetic data generation, model distillation, and high-complexity reasoning.
Parameters: 405 Billion Dense | Context: 128,000 tokens | License: Llama 3.1 Community License
Official DocumentationDeepSeek-V3 (DeepSeek)
DeepSeek-V3 uses Multi-head Latent Attention (MLA) to compress KV caches by 93% and DeepSeekMoE auxiliary-loss-free load balancing. It established unprecedented training cost efficiency while competing at the highest tier of open foundation models.
Parameters: 671B MoE (37B active) | Context: 128,000 tokens | License: MIT
Official DocumentationQwen 2.5 72B Instruct (Alibaba Cloud)
Alibaba's flagship open-weights model trained on 18 trillion tokens across 29 languages. Qwen 2.5 72B demonstrates exceptional performance in mathematical problem solving, schema parsing, and multilingual instruction following.
Parameters: 72.7 Billion Dense | Context: 128,000 tokens (up to 8k output) | License: Apache 2.0
Official DocumentationMistral Large 2 (2407) (Mistral AI)
Mistral Large 2 delivers enterprise-grade code generation, mathematical reasoning, and native fluency in French, German, Spanish, Italian, and 80+ coding languages, designed with strict guardrails against false confidence.
Parameters: 123 Billion Dense | Context: 128,000 tokens | License: Mistral Research License / Commercial API
Official DocumentationGrok 2 (xAI)
Developed by xAI, Grok 2 features advanced vision comprehension, real-time social knowledge synthesis, and conversational agility, paired with FLUX image generation capabilities on the X platform.
Parameters: Undisclosed | Context: 128,000 tokens | License: Proprietary Commercial API
Official DocumentationQwen 2.5 Coder 32B Instruct (Alibaba Cloud)
Trained on 5.5 trillion code tokens, Qwen 2.5 Coder 32B sets a record for open-source programming capabilities. It serves as an ideal local engine for IDE assistants, repository-scale refactoring, and multi-file debugging.
Parameters: 32.5 Billion Dense | Context: 128,000 tokens | License: Apache 2.0
Official DocumentationCodestral 2501 (Mistral AI)
Codestral is specifically tuned for fill-in-the-middle (FIM) code completion, test generation, and documentation. Its expanded 256k context allows it to ingest broad codebase contexts inside VS Code and JetBrains IDE extensions.
Parameters: 22 Billion Dense | Context: 256,000 tokens | License: Mistral Non-Production License / Commercial API
Official DocumentationDeepSeek-Coder-V2 (DeepSeek)
DeepSeek-Coder-V2 was the first open-source model to attain coding scores comparable to GPT-4 Turbo. It supports 338 programming languages and expands context length to 128k tokens for repo-wide code comprehension.
Parameters: 236B MoE (21B active) | Context: 128,000 tokens | License: MIT
Official DocumentationFLUX.1 [dev] (Black Forest Labs)
Created by the original engineers of Stable Diffusion, FLUX.1 leverages 12 billion parameters and flow matching to produce unmatched typographic accuracy, intricate anatomical fidelity, and photorealistic lighting.
Parameters: 12 Billion | Context: Text-to-Image Diffusion | License: Non-Commercial FLUX.1 [dev] License
Official DocumentationStable Diffusion 3.5 Large (Stability AI)
Stability AI's flagship 8.1B parameter open model combines separate text and image weight spaces to allow deeper conceptual alignment, customizable visual styling, and high-resolution rendering without external upscalers.
Parameters: 8.1 Billion | Context: Text-to-Image Diffusion | License: Stability AI Community License
Official DocumentationMidjourney v6.1 (Midjourney)
Midjourney v6.1 improves skin textures, 8-pixel fine details, reduced pixel artifacts, and 25% faster generation speeds, maintaining its status as the premier commercial image generator for digital creators and art directors.
Parameters: Undisclosed | Context: Proprietary Image Synthesis | License: Commercial Subscription
Official DocumentationGoogle Veo 2 (Google DeepMind)
Veo 2 understands cinematic terms such as timelapses, aerial tracking shots, and lens focal lengths. It produces fluid physical simulations, realistic human movement, and temporal consistency across multiple continuous seconds.
Parameters: Undisclosed | Context: High-Definition Video Generation | License: Google Cloud / VideoFX API
Official DocumentationOpenAI Sora (OpenAI)
Sora operates as a diffusion model over spacetime latent patches, generating video by gradually predicting clear frames from visual noise. It maintains continuous object persistence even when subjects temporarily leave the viewport.
Parameters: Undisclosed | Context: Video Generation Engine | License: Commercial Subscription / API
Official DocumentationRunway Gen-3 Alpha (Runway)
Trained jointly on videos and images, Gen-3 Alpha represents Runway's next-generation architecture for Hollywood VFX, commercial advertising, and interactive video synthesis with expressive human acting.
Parameters: Undisclosed | Context: Text/Image to Video | License: Commercial Subscription / API
Official DocumentationWhisper Large v3 Turbo (OpenAI)
Whisper Large v3 Turbo reduces decoder layers from 32 down to 4 while preserving state-of-the-art multilingual word error rates (WER), making real-time open speech-to-text inference feasible on consumer hardware.
Parameters: 809 Million | Context: 30-second sliding audio chunks | License: MIT
Official DocumentationElevenLabs Multilingual v2 (ElevenLabs)
ElevenLabs synthesizes realistic human voices with natural cadence, pauses, emotional inflections, and zero-shot voice cloning capabilities, powering conversational AI agents and audiobook productions worldwide.
Parameters: Undisclosed | Context: Real-time Text-to-Speech | License: Commercial API
Official DocumentationQwen2-Audio 7B (Alibaba Cloud)
Qwen2-Audio can directly accept audio speech, music, or environmental sounds and carry out conversations via text, bypassing errors introduced by intermediate speech-to-text transcription layers.
Parameters: 8.2 Billion | Context: Direct Audio Stream Input | License: Apache 2.0
Official DocumentationQwen 2.5-VL 72B Instruct (Alibaba Cloud)
Qwen 2.5-VL parses arbitrary image resolutions and multi-hour video feeds natively. It achieves state-of-the-art results for document visual QA, OCR, phone/PC GUI bounding box coordinates, and visual math reasoning.
Parameters: 72.5 Billion | Context: 128,000 tokens (Native Dynamic Resolution) | License: Apache 2.0
Official DocumentationPixtral Large (124B) (Mistral AI)
Built upon Mistral Large 2 with a custom 1B parameter vision encoder, Pixtral Large processes arbitrarily sized images and documents without resizing distortion, providing top-tier visual reasoning and chart analysis.
Parameters: 124 Billion Dense | Context: 128,000 tokens | License: Mistral Research License / Commercial API
Official DocumentationGemma 2 27B (Google DeepMind)
Google's Gemma 2 27B delivers competitive performance against models more than twice its size. It utilizes local-global sliding window attention and logit soft-capping for high inference efficiency on modern workstation GPUs.
Parameters: 27.2 Billion Dense | Context: 8,192 tokens | License: Gemma Terms of Use (Open Weights)
Official DocumentationCommand R+ (08-2024) (Cohere)
Command R+ specializes in verifiable enterprise workflows, producing precise citations linked to source documents to eliminate hallucination risk in regulated banking, legal, and operational environments.
Parameters: 104 Billion Dense | Context: 128,000 tokens | License: Commercial API / Non-commercial Research
Official Documentation