🖥️ Локальные LLM

Dense LLM from IBM supporting up to 128K context length, trained on 12T tokens. Suitable for general instructions following and can be used to build AI assistants.

State-of-the-art image + text input models from Google, built from the same research and tech used to create the Gemini models.

The mid-sized option of the Gemma 2 model family. Built by Google, using from the same research and technology used to create the Gemini models.

The large option of the Gemma 2 model family. Built by Google, using from the same research and technology used to create the Gemini models.

Devstral by MistralAI is based on Mistral Small 3.1. Debuts as the #1 open source model on SWE-bench.

A 7B Vision Language Model (VLM) from the Qwen2.5 family.

Advanced open-weight reasoning model, finetuned from Phi-4 with additional reinforcement learning for higher accuracy.

Lightweight open model from the Phi-4 family.

State-of-the-art image + text input models from Google, built from the same research and tech used to create the Gemini models.

State-of-the-art image + text input models from Google, built from the same research and tech used to create the Gemini models. Smallest model in the Gemma 3 family — runs anywhere.

Meta's lightweight multilingual model. Text-only, great for on-device and edge deployment. Supports 8 languages.

Meta's flagship 8B. Multilingual, strong reasoning, tool use. The most popular open model for general tasks.

Meta's most powerful open model. Near GPT-4 class on reasoning, coding, and instruction following.

Mistral AI's groundbreaking 7B. Outperforms larger models on reasoning. Excellent for fine-tuning.

Mistral's Mixture-of-Experts. 8 experts x 7B, activates 2 per token. 46B performance at 12B inference cost.

DeepSeek's reasoning model with chain-of-thought. Shows thinking process. Strong on math, code, logic.

DeepSeek-R1 distilled to 32B (Qwen base). Strong reasoning with visible chain-of-thought. Near GPT-4o on math.

MoE code model from DeepSeek. 338 languages, 128K context. Strongest open code model for single-GPU.

Alibaba's Qwen2.5 7B. 29 languages, 128K context. Great all-rounder for Asian and European languages.

Alibaba's code-specialized model. Trained on 5.5T tokens of code. 92 programming languages.

Meta's code-specialized Llama. Python specialist with fill-in-the-middle. Code completion and generation.

Large Language and Vision Assistant. Llama 2 + vision encoder. Image understanding, OCR, visual QA.

Uncensored Mixtral fine-tune. Removes refusal behaviour. Good for creative and unrestricted tasks.

HuggingFace's DPO-trained Mistral 7B. Punches above weight on chat benchmarks. Excellent conversationalist.

C-RLFT trained Mistral 7B. Top performer on MT-Bench among 7B models. Natural conversation.

01.AI's 34B model. 3T training tokens. Strong bilingual (EN/ZH). Near GPT-3.5 performance.

Cohere's flagship. RAG-optimised, tool-use native, 10 languages. Enterprise-grade reasoning.

BigCode project. The Stack v2 (600+ languages). Fill-in-the-middle, code completion.

Defog.ai's SQL specialist. 20K+ SQL queries across diverse schemas. Text-to-SQL near GPT-4 accuracy.

Microsoft's MoE model. Evolved training on complex instructions. Top reasoning, multilingual, coding.

Nous Research fine-tune on 1M+ instructions. Structured output, function calling, JSON mode. ChatML format.

Databricks' MoE flagship. 16 experts, 4 active. Strong at structured data, SQL, Python, reasoning.

OpenBMB's vision model. Strong OCR, image understanding, bilingual (EN/ZH). Runs on edge devices.

TII's latest Falcon. 14T training tokens. Strong multilingual (EN, FR, ES, PT). Function calling, code.

Nomic AI's embedding model. #1 on MTEB among open models. 768-dim vectors. Perfect for RAG and search.

Mixedbread AI's embedding. 1024-dim vectors, MTEB leader. Best open embedding for retrieval tasks.