Skip to content

Foundation_Models.md

Latest commit

 

History

History
417 lines (299 loc) · 45 KB

Foundation_Models.md

File metadata and controls

417 lines (299 loc) · 45 KB
Raw

Foundation Models

Large language models and multimodal models. This file is organized into language-only LLMs (general-purpose models, domain fine-tuned variants, and the providers and engines used to run them) and multimodal models (vision-language and omni models).

Fast-moving general rankings are delegated to external leaderboards that stay current. The curated tables here cover the niche, fine-tuned, and infrastructure tools that those leaderboards do not track.


Table of contents


Language Only Large Language Models

Large language models (LLMs) are systems trained on large text corpora to recognize patterns and generate natural language. This section covers instruction-tuned general-purpose models, domain fine-tuned variants, and the providers and engines used to run them.

Note

Effective use of any model depends on understanding its capabilities, its limitations, and the strategies needed to mitigate bias.

Advanced Language and Reasoning LLMs

General-purpose model quality changes from week to week, so a hand-maintained ranking goes stale quickly. For current comparisons across quality, context window, speed, and price, use the leaderboards below. Each links to a filtered view; on Artificial Analysis you can add or remove models to build your own comparison.

Note

Open-source models

Proprietary models

Tip

To run open-source models on your own hardware, see How to run LLMs on your machine.



Finetuned LLMs

Fine-tuned LLMs are models adapted to a particular domain, task, or dataset. This adaptation typically improves performance and accuracy within that context, compared to general-purpose use. The categories below focus on axes that general-purpose leaderboards do not rank: language-specific performance, role-play, and uncensored behavior.

Language Specific

Models fine-tuned for a specific language or language family.

Indian Indian languages

Organization Model Name Model Sizes Context Window Reasoning Model
Sarvam AI Sarvam-m 24B 131k ✔️

japan Japanese

Organization Model Name Model Sizes Context Window Reasoning Model
Shisa Shisa-v2-llama3.1-405b 405B 128k ❌
Shisa Shisa-v2-llama3.3-70b 70B 128k ❌
Shisa Shisa-v2-qwen2.5-32b 32B 128k ❌
Shisa Shisa-v2-mistral-small-24b 24B 128k ❌
Shisa Shisa-v2-mistral-nemo-12b 12B 128k ❌
Shisa Shisa-v2-mistral-llama3.1-8b 8B 128k ❌
Shisa Shisa-v2-mistral-qwen2.5-7b 7B 128k ❌
ELYZA Llama-3-ELYZA-JP-8B 8B 8k ❌
Swallow Llama-3.1-Swallow-8B-Instruct-v0.3 8B 128k ❌

korea Korean

Organization Model Name Model Sizes Context Window Reasoning Model
sktelecom A.X-4.0 72B 128k ❌
sktelecom A.X-4.0-Light 7B 131k ❌
kt Midm-2.0-Base-Instruct 12B 32k ❌
kt Midm-2.0-Mini-Instruct 2B 32k ❌
Naver HyperCLOVAX-SEED-Think-14B 14B ? ✔️
Trillion Labs Trillion-7B-preview 7B ? ❌

Arabic Arabic

Organization Model Name Model Sizes Context Window Reasoning Model
Inception Jais-family-30B-16K-chat 30B 16k ❌
ALLaM ALLaM-7B-Instruct 7B ? ❌

Southeast Asian Southeast Asian

Organization Model Name Model Sizes Context Window Reasoning Model
Sea AI Lab Sailor2-20B-Chat 20B 32k ❌
AI Singapore Llama-SEA-LION-v3-8B-IT 8B 128k ❌

European European

Organization Model Name Model Sizes Context Window Reasoning Model
UTTER EuroLLM-9B-Instruct 9B 4k ❌
BSC Salamandra-7B-Instruct 7B 8k ❌

Role Play

Models optimized for role-play tasks through specialized training data.

Role-playing is a technique where the model assumes a specific character, profession, or persona to generate more focused, contextually relevant responses.

Large-scale models (70+ billion parameters) : These require substantial amounts of RAM and GPU memory, making local installation impractical for most users. As a result, these models are typically deployed on cloud-based platforms that provide the necessary computational resources.

Organization Base Model Finetuned Model Model Sizes Context Window Reasoning Model Geographic Origin
Steelskull Llama-3.3-70B-Instruct L3.3-Electra-R1-70b 70B 128K ❌ usa
Steelskull Llama-3.3-70B-Instruct L3.3-MS-Nevoria-70b 70B 128K ❌ usa
LatitudeGames Llama-3.3-70B-Instruct Wayfarer-Large-70B-Llama-3.3 70B 128K ❌ usa
BosonAI Llama-3-70B-Instruct Higgs-Llama-3-70B 70B 32K ❌ usa
TheDrummer Llama-3.3-70B-Instruct Anubis-70B-v1.1 70B 128K ❌ usa
Sao10K Llama-3.3-70B-Instruct L3.3-70B-Euryale-v2.3 70B 128K ❌ usa
Anthracite Qwen2.5-72B-Instruct magnum-v4-72b 72B 128K ❌ China

Mid-sized and smaller models (less than 70 billion parameters) : These models are suitable for local deployment on high-end to mid-range consumer setup. However, such deployments require a significant hardware investment, including a GPU (16–32 GB of VRAM) and related components, typically resulting in total costs ranging from $1000 to $3,000 (or equivalent).

Organization Base Model Finetuned Model Model Sizes Context Window Reasoning Model Geographic Origin
Gryphe Qwen3-30B-A3B-Base Pantheon-Proto-RP-1.8-30B-A3B 30B 128K Hybrid China
LatitudeGames Mistral-Small-3.1-24B-Instruct-2503 Harbinger-24B 24B 128K ❌ eu
Gryphe Mistral-Small-3.1-24B-Instruct-2503 Pantheon-RP-1.8-24b-Small-3.1 24B 128K ❌ eu
Gryphe Mistral-Small-Instruct-2409 Pantheon-RP-Pure-1.6.2-22b-Small 22B 128K ❌ eu
LatitudeGames Mistral-Nemo-Instruct-2407 Muse-12B 12B 128K ❌ eu
LatitudeGames Mistral-Nemo-Instruct-2407 Wayfarer-12B 12B 128K ❌ eu
Wave Mistral-Nemo-Base-2407 silly-v0.2 12.2B 128K ❌ eu
Oxygen Qwen2.5-14B-Instruct Oxy-1-small 14.8B 131K ❌ China
Oxygen Qwen2-1.5B Oxy-1-micro 1.54B 131K ❌ China
TheDrummer Mistral-Small-24B-Instruct-2501 Cydonia-24B-v2 24B 32K ❌ eu
TheDrummer Mistral-Nemo-Instruct-2407 Rocinante-12B-v1.1 12B 128K ❌ eu


Uncensored

Mainstream leaderboards do not measure how readily a model answers contentious or restricted prompts, so this ranking is delegated to a leaderboard built for that axis.

Note



LLM Providers

Cloud-based LLM Providers

Tip

Reference the Artificial Analysis Leaderboard for comparisons of LLM providers across key metrics: pricing, token generation speed, response latency, and context window.

Major Cloud AI Providers

These are typically large tech companies that develop and offer access to their own flagship LLMs. Access is usually via APIs, and they manage all the underlying infrastructure.

Tool Description Models Pricing
AI21 Labs Develops the Jamba family of hybrid SSM-Transformer models, focused on long context and efficiency. Jamba Large 1.6 Freemium
Anthropic Develops the Claude family of models, trained with Constitutional AI to be helpful, harmless, and honest. Claude Fable5, Claude Opus 4.8 and more Freemium
Cohere Develops the Command family of enterprise-focused models for retrieval, generation, and tool use. Command A Freemium
Databricks A unified, open analytics platform that provides tools and services for data processing, analytics, and artificial intelligence at scale. Dbrx-instruct Paid
Deepseek A Chinese AI lab known for its open-weight DeepSeek models, with strong reasoning at low cost. DeepSeek-V4-Pro and more Freemium
Google Develops the Gemini family of multimodal models for language, vision, and reasoning. Gemini 3.5, Gemini 3.5 Flash and more Freemium
Mistral AI A French AI company developing open-weight and commercial language models, agents, and developer tools. Mistral Medium 3.5 and more Freemium
OpenAI Develops the GPT family of language models, along with DALL-E for images and Whisper for speech. GPT-5.5 and more Freemium
Qwen Alibaba's Qwen model family, spanning open-weight and hosted models. Qwen3.7-Max and more Freemium
Reka Develops multimodal Reka models with strong vision, audio, and video understanding. Reka Flash 3.1 and more Freemium
xAI Develops the Grok family of models, integrated with the X platform and available via API. Grok 4.3 and more Freemium


Aggregator & Specialized Cloud Platforms

These platforms offer access to a variety of models from different developers, including many popular open-source models and sometimes proprietary ones, often through a single API or interface. Some may offer specialized hardware or value-added services like fine-tuning.

Tool Description Models Pricing
Amazon Web Services (AWS) Provides hosted access to many open-source and proprietary models through Amazon Bedrock and SageMaker. Large Panel of Open source and Proprietary Models Paid
Baseten An inference platform for deploying and serving open-source and custom models in production, with autoscaling and low-latency serving. Large Panel of Open source Models Paid
Cerebras An AI company that has developed innovative hardware and software solutions for AI computing. Llama-3.3-70B-Instruct and more Freemium
DeepInfra A platform that provides scalable and cost-effective infrastructure for deploying machine learning models. Large Panel of Open source Models Freemium
Featherless A subscription-based, serverless LLM hosting provider that gives you instant, unlimited access to thousands of Hugging Face models, all via a simple API and with no infrastructure hassles. Large Panel of Open source Models Freemium
Fireworks An inference platform for deploying and fine-tuning open-source models in production. Large Panel of Open source Models Freemium
Groq Specializes in high-performance AI inference with custom LPU (Language Processing Unit) hardware, offering models like Meta's Llama 3. Llama-3.3-70B-Instruct and more Freemium
Hyperbolic An open-access AI cloud platform offering affordable GPU compute and open-source model inference. Large Panel of Open source Models Freemium
LeptonAI A platform that provides cloud-based infrastructure and tools for deploying and running AI applications efficiently. Large Panel of Open source Models Freemium
Lumo by Proton A privacy-first AI assistant developed by Proton, the company known for Proton Mail, Proton VPN, and Proton Pass. Panel of Open source Models Freemium
Microsoft Azure A comprehensive suite of AI services and tools designed to help developers and organizations build, deploy, and manage AI applications at scale. Large Panel of Open source and Proprietary Models Paid
Nebius An inference-as-a-service platform for running open-source models Large Panel of Open source Models Freemium
Novita An inference-as-a-service platform for running open-source models Large Panel of Open source Models Freemium
OpenRouter A versatile platform designed to provide access to a wide range of large language models (LLMs) from both proprietary and open-source sources. Large Panel of Open source and Proprietary Models Paid
Parasail An AI infrastructure platform for deploying and scaling open-source models. Large Panel of Open source Models Freemium
Poe An AI chatbot aggregator platform developed by Quora that provides users access to multiple advanced language models and chatbots within a single interface. Large Panel of Open source and Proprietary Models Freemium
Replicate A platform for running and deploying open-source models through a simple API; now part of Cloudflare. Large Panel of Open source Models Paid
SambaNova An artificial intelligence company that provides a comprehensive AI platform for enterprises. Llama-3.3-70B-Instruct and more Paid
Together A cloud platform designed for building and running generative AI applications. Large Panel of Open source Models Paid


Local LLM Providers

Important

To run LLMs locally, see the setup guide for private, on-device model use.

Tool Description OS Models
AnythingLLM An open-source, full-stack application that allows users to chat with their documents in a private and enterprise-friendly environment. All All Open source Models
Chatbox AI-powered conversational interface that enables human-like interactions through text or voice. All All Open source Models
ChatWise A high-performance, privacy-focused AI chatbot platform that supports multiple LLMs for versatile, multimodal interactions. All Large Panel of Models
Cherry Studio A cross-platform desktop application that serves as a unified interface for interacting with multiple large language models (LLMs)—both cloud-based and locally hosted. All Large Panel of Models
Enchanted iOS and macOS app for chatting with private self hosted language models. MacOS/IOS Large Panel of Open source Models
FreeChat An AI-powered chat application designed specifically for macOS. MacOS Large Panel of Open source Models
Google AI Edge Gallery Open-source app designed to showcase and let users interact with on-device machine learning (ML) and generative AI (GenAI) models directly on their smartphones. Android / iOS(Soon) All Open source Models
GPT4ALL An open-source software ecosystem developed by Nomic AI that enables users to run powerful large language models (LLMs) locally on their personal computers. All Large Panel of Open source Models
Jan Clean UI with useful features like system monitoring and LLM library. All Large Panel of Open source Models
LibreChat Open-source chat interface that supports multiple AI models, including Anthropic, AWS, OpenAI, and Azure. It offers features like agents with file handling, a code interpreter for various languages. All Large Panel of Models
LLMFarm An open-source app for running LLMs locally and offline on iOS and macOS devices. iOS/ MacOS All Open source Models
LM Studio Elegant UI with the ability to run every Hugging Face repository. All Large Panel of Open source Models
Msty An AI chat application that offers a user-friendly interface for interacting with both local and online AI language models. All Large Panel of Open source Models
Ollama Fastest when used on the terminal, and any model can be downloaded with a single command. All All Open source Models
Open WebUI Self-hosted, open-source web interface designed for running and managing LLMs locally or offline. All All Open source Models
Silly Tavern Open-source LLM frontend designed for power users. All All Open source Models
Witsy Open-source LLM frontend designed for power users. All All Open source Models
Pocketpal AI assistant powered by small language models (SLMs) that run directly on your phone. Android / iOS All Open source Models


LLM Inference Engines

This table lists notable frameworks and libraries designed for running and serving Large Language Models efficiently, focusing on performance, scalability, and deployment.

The landscape evolves rapidly. Some tools might integrate techniques or components from others (e.g., many serving frameworks might leverage optimized kernels like FlashAttention or specific backend libraries like vLLM or TensorRT-LLM). This table focuses on the primary offering or framework level.

Engine/Server Developer/Origin Key Features & Focus Primary Use Case(s) Notes
llama.cpp Georgi Gerganov et al. Highly optimized C++ inference for GGUF models; CPU & GPU (Metal, CUDA, OpenCL) support. Local inference, experimentation, backend for other tools Foundational library/CLI, broad hardware support.
MLX LM Apple Inference and fine-tuning on Apple Silicon; built on the MLX unified-memory framework. Local inference on Mac Powers Ollama's Apple Silicon backend as of 2026.
ExLlamaV2 turboderp Fast inference for NVIDIA consumer GPUs; EXL2 quantization format and custom CUDA kernels. Local inference (consumer NVIDIA GPUs) Among the fastest single-user engines on RTX hardware.
vLLM vLLM Project (Berkeley) High-throughput serving library; PagedAttention, continuous batching. Production serving, research Primarily Python library, integrates with frameworks like Ray, OpenLLM.
Text Generation Inference (TGI) Hugging Face Production-ready server; Optimized for HF models, high throughput, streaming. Production serving Rust/Python based; archived and in maintenance mode since 2026, with Hugging Face pointing users to vLLM and SGLang.
TensorRT-LLM NVIDIA Optimization library & engine for NVIDIA GPUs; Quantization, in-flight batching. High-performance serving (NVIDIA HW) Requires model compilation, integrates with Triton.
Triton Inference Server NVIDIA General-purpose inference server; Supports multiple frameworks (TRT, PyTorch, TF). Production serving (diverse models) Can serve LLMs (often via TensorRT-LLM backend) and other models.
OpenLLM BentoML Production-ready LLM serving framework; Integrates vLLM/BentoML, OpenAI API compat. Production serving Aims to simplify deployment and scaling.
Ray Serve Anyscale (Ray Team) Scalable model serving library on Ray; Distributed computing support. Distributed production serving General purpose, but powerful for scaling LLM endpoints.
LMDeploy OpenMMLab / MMDeploy Efficient inference framework; Quantization, TurboMind engine. Research, production serving Part of the OpenMMLab ecosystem.
MLC LLM MLC AI Team Universal deployment solution; Compiles models for diverse hardware (CPU, GPU, mobile). Cross-platform deployment Focuses on Machine Learning Compilation.
SGLang SGL Project Efficient structured generation & inference; RadixAttention. Research, production serving Optimized for complex generation tasks.
DeepSpeed Inference Microsoft Optimized inference kernels & engine (part of DeepSpeed library). High-performance serving Leverages techniques developed for large-scale training.
Xinference Xorbits Platform for deploying LLMs & embedding models; API compatible with popular tools. Local & cloud deployment Aims to be a unified deployment solution.
LocalAI Go-Skynet OpenAI-compatible API layer; Pluggable backends (incl. llama.cpp, but others too). Local development, API replacement Acts as a bridge, not the core engine itself.
LLM Engine (Scale) Scale AI Fine-tuning & serving platform; Optimized inference, simplified deployment. Production serving (hosted or self-hosted) Provides both infrastructure and optimization.


Multimodal Language Models


Vision Language Models

Vision Language Models (VLMs) are multimodal models that understand and generate content by combining visual and text input, supporting tasks such as image captioning, visual question answering, and object recognition. For technical background, see the Hugging Face VLM overview.

The VLM landscape moves as quickly as the text-only one, so the size-tiered ranking that previously lived here is delegated to external leaderboards that are kept current.

Note

Tip

To run open-source VLMs on your own hardware, see How to run LLMs on your machine.



Omni Large Language Models

Omni Large Language Models (OLLMs) extend beyond image and text to handle audio, video, and real-time streaming data, supporting tasks such as video transcription, audiovisual reasoning, and multimodal integration. They can produce text, audio, and image output.

This is a fast-moving, mostly open-weight niche with no established leaderboard, so the table below is a curated snapshot of current open models rather than a ranking.

Open-Source OLLMs

Organization Model Name Model Sizes Input Output Geographic Origin
Alibaba Qwen3-Omni-30B-A3B-Instruct 30B (A3B) Text, Image, Audio, Video Text, Speech China
OpenBMB MiniCPM-o 4.5 9B Text, Image, Audio, Video Text, Speech China
AntGroup Ming-Lite-Omni-1.5 19.1B Text, Image, Audio, Video Text, Image, Speech China
Baichuan Baichuan-Omni-1.5 7B Text, Image, Audio, Video Text, Speech China
VITA VITA-1.5 7B Text, Image, Audio, Video Text, Speech China
Infinigence Megrez-3B-Omni 3B Text, Image, Audio Text China
Deepseek Janus-Pro-7B 7B Text, Image Text, Image China
flagopen Emu3 8B Text, Image, Video Text, Image, Video China
ByteDance BAGEL-7B-MoT 7B (14B MoE) Text, Image Text, Image China