متطلبات الأجهزة لنماذج الذكاء الاصطناعي مفتوحة المصدر Hardware Requirements
قارن متطلبات الأجهزة لنماذج الذكاء الاصطناعي الشائعة مفتوحة المصدر، واعثر على حجم النموذج المناسب لخادمك.
DeepSeek Models
Keep compact R1 distills for realistic home deployment; keep V4 to answer when the full model moves into server-class hardware.
Qwen Models
Qwen3 provides one of the cleanest hardware ladders from sub-2B local models to workstation and server-class MoE deployments.
Gemma Models
Gemma 4 is especially useful for this catalog because Google publishes size-specific local and edge memory guidance.
Llama Models
Keep lightweight Llama 3.x models for realistic local use and Llama 4 only where the hardware question itself is valuable.
Mistral Models
Prioritize current open models that span edge, coding workstation, and server-class deployment rather than preserving every historical Mistral release.
Phi Models
Small first-party models with strong edge and local-inference value; behavior variants with the same hardware envelope should stay inside one hardware page.
GPT-OSS Models
OpenAI's open-weight reasoning models are included because they explicitly target local deployment—from 16 GB-class systems to 80 GB accelerator hardware.
Kimi Models
Kimi stays in the catalog for hardware-decision value, not because compact Zima systems are the right target for the full frontier checkpoints.
GLM Models
Mix one practical smaller GLM profile with flagship models whose main value is showing where home hardware stops being a sensible recommendation.
Nemotron Models
Keep only releases with clear local/accelerated deployment value instead of mirroring NVIDIA's full model library.
Granite Models
Granite 4.1 is a strong Zima-oriented family because IBM ships clear 3B, 8B, and 30B hardware tiers and documents local Ollama/LM Studio workflows.
OLMo Models
Use hardware-profile pages by parameter class; Base, Instruct, and Think variants can be explained inside the same 7B or 32B hardware guide when their memory envelope is materially similar.
