Private language models
Compare self-hosted language models for assistants, analysis, long documents, structured output and tool use. Hardware fit starts with the exact checkpoint and precision, then expands for context, cache and concurrent users.
Selection approach
Start with the smallest model that passes your own accuracy and reasoning tests. Larger total parameter counts can improve capability, but sparse activation, runtime maturity and memory format materially change cost and speed.
01
02
03
17 current models
Showing 17 models
Moonshot AI
A 2.8-trillion-parameter, 104-billion-active multimodal mixture-of-experts model for long-context reasoning, coding, visual understanding and agentic work.
Qwen
A 36-billion-parameter multimodal mixture-of-experts model with about 3 billion active parameters, a 262,144-token default context and a strong emphasis on agentic coding.
DeepSeek
A 284B-total, 13B-active mixture-of-experts language model with one-million-token context and a mixed FP4/FP8 release format.
Mistral AI
A 119B-total, 6B-active hybrid model that combines instruction following, reasoning, coding-agent and multimodal capabilities.
Meta
Meta's 109B-total, 17B-active multimodal mixture-of-experts checkpoint with text and image input and a very long documented context.
OpenAI
OpenAI's 117B-total, 5.1B-active open-weight reasoning and agentic model, released with native MXFP4 MoE weights.
Google's 27B instruction-tuned multimodal model for text and image input, with a 128K context and broad language coverage.
OpenAI
A 21-billion-parameter sparse reasoning model with 3.6 billion active parameters, native tool use and an MXFP4 release intended for local or specialised work.
Qwen
An 80-billion-total, 3-billion-active sparse model designed for coding agents, local development and long repository context.
Qwen
A compact 4-billion-parameter multimodal model with native vision, reasoning, coding and broad multilingual support.
Qwen
A 27-billion-parameter dense multimodal model for reasoning, coding, agents and visual understanding across 201 languages and dialects.
Google DeepMind
Google DeepMind's 12-billion-parameter instruction-tuned Gemma 4 model, supporting text, images, video and audio input with text output.
Google DeepMind
Google DeepMind's 30.7-billion-parameter instruction-tuned multimodal model for text and image understanding with a 256K context window.
StepFun
A 198-billion-parameter sparse vision-language model with about 11 billion active parameters, designed for reasoning, coding, agents and visual intelligence.
Z.ai
A 744-billion-total, 40-billion-active sparse model aimed at complex systems engineering and long-horizon agentic tasks.
MiniMax
A sparse agentic model trained for coding, tool use, search and office work across more than ten programming languages.
StepFun
An end-to-end audio-language model for spoken conversation, audio understanding, paralinguistic cues and audio tool use.
No model matches that search and workload combination.
Model-to-hardware fit
Use the precise model version, numerical format and complete software pipeline intended for production.
Check whether it can load on one GPU or requires supported multi-GPU loading.
Allow for context, cache, batch, media encoders and concurrent users.
Measure quality, latency and stability on representative work.
Model weights are only one part of the system. Data access, runtime software, storage, monitoring and administrator handover also affect a reliable deployment.
Privacy & cookies. Google Analytics measures site use so we can improve it. Read our Cookie Notice and Privacy Notice. To object, disable cookies in your browser.