Summary

Okay, here's a breakdown of the provided model data, categorized for easier understanding:

I. Overview - Model Size & Architecture

* Dense Models: These models have a relatively uniform architecture – essentially a large number of interconnected neurons.
* Smaller Dense Models (under 30B parameters):
* Qwen 2.5: Ranges from 32B to 72B parameters. Notable for multilingual capabilities and reasoning.
* Command R: 35B parameters, specialized for retrieval-augmented generation (RAG).
* OLMo 2: 32B parameters, focuses on research with a smaller context window.
* EXAONE 4.0: 32B parameters, leveraging hybrid reasoning.
* Larger Dense Models (30B+ parameters):
* Qwen 3.5: Ranges from 122B to 397B parameters. Significant improvements in quality and multimodal capabilities (vision, reasoning, code).
* Qwen 3: 235B and 480B parameters, featuring large MoE architectures.
* Mixtral 8x7B: 47B parameters. A Mixture of Experts model, meaning it utilizes multiple smaller networks that are activated based on the input.
* Llama 4 Maverick: 405B parameters.
* Mixture of Experts (MoE) Models: These models utilize a collection of smaller, specialized networks ("experts"). Only a subset of these experts is activated for any given input, leading to increased efficiency.
* Mixtral: Available in 8x7B and 8x22B variants. The "x" indicates the number of experts.
* Qwen 3: Various sizes (122B, 235B, 397B, 480B) all utilize MoE.
* Llama 4 Maverick: Utilizes a 128E architecture.
* Qwen 3.5: Various sizes (122B, 397B) all utilize MoE.
* DeepSeek V3.1: 671B uses a hybrid thinking and tool use architecture.
* DeepSeek:
* DeepSeek V3.1: 671B, with a hybrid approach to thinking and tool use.

II. Key Features & Capabilities

* Multilingual: Several models (Qwen 2.5, Qwen 3.5, Qwen 3) are specifically designed for multilingual performance.
* Reasoning: Many of the larger models (Qwen 3.5, Qwen 3, Llama 4, DeepSeek V3.1) demonstrate strong reasoning abilities.
* Code Generation: Qwen 3, Qwen 3.5, and the Qwen 2.5 models are optimized for code generation.
* Multimodal: Qwen 3.5 and Qwen 3 are multimodal, capable of processing both text and image inputs.
* Retrieval-Augmented Generation (RAG): Command R is optimized for this.
* Tool Use: DeepSeek V3.1 improves on tool use.

III. Model Size Comparison

| Model Name | Parameters (Approx.) | Context Window |
| ------------------------ | -------------------- | -------------- |
| Qwen 2.5 (32B) | 32B | 2048 |
| Qwen 2.5 (72B) | 72B | 2048 |
| Command R (35B) | 35B | 2048 |
| OLMo 2 (32B) | 32B | 4K |
| Mixtral 8x7B | 47B | 2048 |
| Mixtral 8x22B | 141B | 2048 |
| Qwen 3 (122B) | 122B | 2048 |
| Qwen 3 (235B) | 235B | 2048 |
| Qwen 3 (397B) | 397B | 2048 |
| Qwen 3 (480B) | 480B | 2048 |
| Qwen 3.5 (122B) | 122B | 2048 |
| Qwen 3.5 (397B) | 397B | 2048 |
| Llama 4 Maverick (405B) | 405B | 128K |
| Qwen 3 Coder (480B) | 480B | 256K |
| DeepSeek V3.1 (671B) | 671B | 128K |

Notes:

* Context Window: Refers to the amount of text the model can consider at once. Larger windows generally allow for more coherent and contextually aware responses.
* "A3B" indicates the use of "active" experts in the Mixture of Experts architecture.
* The information is based on the data provided. Specific details about training data, fine-tuning, and performance metrics would require further investigation.

To help me provide even more focused information, could you tell me:

* What are you specifically interested in knowing about these models? (e.g., comparison of their reasoning abilities, best model for a particular task, etc.)

Title

CanIRun.ai — Can your machine run AI models?

Description

Detect your hardware and find out which AI models you can run locally. GPU, CPU, and RAM analysis in your browser.

Keywords

architecture, memory, chat, reasoning, active, year, code, model, vision, mistral, gemma, llama, open, context, edge, google, month

NS Lookup

A 172.67.201.155, A 104.21.44.173

Dates

Created 2026-03-14

Updated 2026-03-14

Summarized 2026-03-14