Skip to content
AI Compute Radar

Hardware guide

Best local models for
RTX 5090

◇ Estimated— Estimated: computed from our curated model and hardware catalog — not a live reading.

Also in:Español

Specifications

Memory
32 GB
Memory type
Dedicated VRAM
Bandwidth
1792 GB/s
Class
Top consumer

The short answer

Qwen AgentWorld 35B-A3B

Weights ~20.6 GB (Q4_K_M), context ~0.2 GB at 8K tokens, runtime ~1.5 GB. Your 32 GB leaves ~28.2 GB usable after a 12% safety margin.

GGUF (Q4_K_M) ↗ (External link)Original (safetensors) ↗ (External link)

Fit labels

  • Runs easily
  • Runs
  • Barely fits
  • Spills into RAM
  • Too big

Every curated model on this card

Context length: 8K tokens · f16 (default)

Model fit on the selected hardware
ModelFitEst. memoryEstimated memory = weights + context + runtime
Qwen3-Coder 30B-A3BAlibaba Qwen · 30.5B MoE · Q4_K_MGGUF ↗ (External link)Original ↗ (External link)Ollama ↗ (External link)Runs easily~19.4 GB17.3 + 0.8 + 1.4
GLM-4.7-FlashZhipu AI · 31.2B MoE · Q4_K_MGGUF ↗ (External link)Original ↗ (External link)Runs easily~18.9 GB17.1 + 0.4 + 1.4
Gemma 4 26B-A4BGoogle · 25.8B MoE · Q4_K_MGGUF ↗ (External link)Original ↗ (External link)Ollama ↗ (External link)Runs easily~17.5 GB15.8 + 0.4 + 1.3
Gemma 3 27BGoogle · 27.4B · Q4_K_MGGUF ↗ (External link)Original ↗ (External link)Ollama ↗ (External link)Runs easily~17.7 GB15.4 + 1.0 + 1.3
Qwen3.8 27BAlibaba Qwen · 27.8B · Q4_K_MGGUF ↗ (External link)Original ↗ (External link)Ollama ↗ (External link)Runs easily~17.1 GB15.3 + 0.5 + 1.3
LFM2.5 24B-A2BLiquid AI · 24B MoE · Q4_K_MGGUF ↗ (External link)Original ↗ (External link)Ollama ↗ (External link)Runs easily~14.9 GB13.5 + 0.2 + 1.3
Mistral Small 3.2Mistral AI · 24B · Q4_K_MGGUF ↗ (External link)Original ↗ (External link)Ollama ↗ (External link)Runs easily~15.8 GB13.3 + 1.3 + 1.2
Devstral Small 2 24BMistral AI · 24B · Q4_K_MGGUF ↗ (External link)Original ↗ (External link)Ollama ↗ (External link)Runs easily~15.8 GB13.3 + 1.3 + 1.2
gpt-oss 20BOpenAI · 20.9B MoE · Q4_K_MGGUF ↗ (External link)Original ↗ (External link)Ollama ↗ (External link)Runs easily~12.2 GB10.8 + 0.2 + 1.2
Qwen3 14BAlibaba Qwen · 14.8B · Q4_K_MGGUF ↗ (External link)Original ↗ (External link)Ollama ↗ (External link)Runs easily~10.8 GB8.4 + 1.3 + 1.1
Gemma 4 12BGoogle · 12B · Q4_K_MGGUF ↗ (External link)Original ↗ (External link)Ollama ↗ (External link)Runs easily~8.1 GB6.6 + 0.4 + 1.0
Nemotron Nano 9B v2NVIDIA · 8.9B · Q4_K_MGGUF ↗ (External link)Original ↗ (External link)Runs easily~8.9 GB6.1 + 1.8 + 1.0
Ornith 1.5 9BOrnith AI · 9.7B · Q4_K_MGGUF ↗ (External link)Original ↗ (External link)Ollama ↗ (External link)Runs easily~6.7 GB5.4 + 0.3 + 1.0
Granite 4.2 8BIBM · 8.8B · Q4_K_MGGUF ↗ (External link)Original ↗ (External link)Ollama ↗ (External link)Runs easily~7.3 GB5.0 + 1.3 + 1.0
LFM2.5 8B-A1BLiquid AI · 8.5B MoE · Q4_K_MGGUF ↗ (External link)Original ↗ (External link)Ollama ↗ (External link)Runs easily~5.9 GB4.8 + 0.1 + 1.0
Qwen3 8BAlibaba Qwen · 8.2B · Q4_K_MGGUF ↗ (External link)Original ↗ (External link)Ollama ↗ (External link)Runs easily~6.8 GB4.7 + 1.1 + 1.0
Gemma 4 E4BGoogle · 8B · Q4_K_MGGUF ↗ (External link)Original ↗ (External link)Ollama ↗ (External link)Runs easily~5.8 GB4.6 + 0.3 + 1.0
Ling 3.0 TinyInclusionAI · 7.9B MoE · Q4_K_MGGUF ↗ (External link)Original ↗ (External link)Runs easily~5.7 GB4.5 + 0.2 + 1.0
Qwen3 4BAlibaba Qwen · 4B · Q4_K_MGGUF ↗ (External link)Original ↗ (External link)Ollama ↗ (External link)Runs easily~4.3 GB2.3 + 1.1 + 0.9
Phi-4 MiniMicrosoft · 3.8B · Q4_K_MGGUF ↗ (External link)Original ↗ (External link)Ollama ↗ (External link)Runs easily~4.2 GB2.3 + 1.0 + 0.9
Granite 4.2 3BIBM · 3.7B · Q4_K_MGGUF ↗ (External link)Original ↗ (External link)Ollama ↗ (External link)Runs easily~3.6 GB2.1 + 0.6 + 0.9
LFM2.5 2.6BLiquid AI · 2.7B · Q4_K_MGGUF ↗ (External link)Original ↗ (External link)Ollama ↗ (External link)Runs easily~2.6 GB1.6 + 0.1 + 0.9
LFM3 1.7BLiquid AI · 1.7B · Q4_K_MGGUF ↗ (External link)Original ↗ (External link)Runs easily~2.0 GB1.0 + 0.1 + 0.9
LFM2.5 1.2BLiquid AI · 1.2B · Q4_K_MGGUF ↗ (External link)Original ↗ (External link)Ollama ↗ (External link)Runs easily~1.7 GB0.7 + 0.1 + 0.9
Qwen3 0.6BAlibaba Qwen · 0.6B · Q8_0GGUF ↗ (External link)Original ↗ (External link)Ollama ↗ (External link)Runs easily~2.3 GB0.6 + 0.9 + 0.9
Qwen AgentWorld 35B-A3BAlibaba Qwen · 34.7B MoE · Q4_K_MGGUF ↗ (External link)Original ↗ (External link)Runs~22.2 GB20.6 + 0.2 + 1.5
Ornith 1.5 35B-A3BOrnith AI · 36B MoE · Q4_K_MGGUF ↗ (External link)Original ↗ (External link)Ollama ↗ (External link)Runs~21.8 GB20.2 + 0.2 + 1.5
KAT-Coder V2.5Kwaipilot · 34.7B MoE · Q4_K_MGGUF ↗ (External link)Original ↗ (External link)Runs~21.5 GB19.9 + 0.2 + 1.4
Qwen3 32BAlibaba Qwen · 32.8B · Q4_K_MGGUF ↗ (External link)Original ↗ (External link)Ollama ↗ (External link)Runs~21.8 GB18.4 + 2.0 + 1.4
Gemma 4 31BGoogle · 31.3B · Q4_K_MGGUF ↗ (External link)Original ↗ (External link)Ollama ↗ (External link)Runs~19.9 GB17.1 + 1.4 + 1.4
Granite 4.2 30BIBM · 29.3B · Q4_K_MGGUF ↗ (External link)Original ↗ (External link)Ollama ↗ (External link)Runs~19.8 GB16.5 + 2.0 + 1.3
Ornith 1.5 122BOrnith AI · 122B · Q4_K_MGGUF ↗ (External link)Original ↗ (External link)Spills into RAM~73.8 GB70.6 + 0.2 + 3.0
Ling 3.0 FlashInclusionAI · 127B MoE · Q4_K_MGGUF ↗ (External link)Original ↗ (External link)Spills into RAM~73.4 GB70.1 + 0.4 + 3.0
GLM-4.5-AirZhipu AI · 110B MoE · Q4_K_MGGUF ↗ (External link)Original ↗ (External link)Spills into RAM~72.3 GB68.0 + 1.4 + 2.9
gpt-oss 120BOpenAI · 117B MoE · Q4_K_MGGUF ↗ (External link)Original ↗ (External link)Ollama ↗ (External link)Spills into RAM~61.4 GB58.5 + 0.3 + 2.6
Qwen3-Coder NextAlibaba Qwen · 79.7B MoE · Q4_K_MGGUF ↗ (External link)Original ↗ (External link)Ollama ↗ (External link)Spills into RAM~48.1 GB45.1 + 0.8 + 2.2
Llama 3.3 70BMeta · 70.6B · Q4_K_MGGUF ↗ (External link)Original ↗ (External link)Ollama ↗ (External link)Spills into RAM~44.1 GB39.6 + 2.5 + 2.0
Qwen3.8 2.4T-A95BAlibaba Qwen · 2.4T-A95B · IQ4_XSGGUF ↗ (External link)Original ↗ (External link)Too big~1259.0 GB1220.8 + 0.7 + 37.5
DeepSeek-V4-ProDeepSeek · 1650B MoE · Q4_KGGUF ↗ (External link)Original ↗ (External link)Too big~816.4 GB791.3 + 0.5 + 24.6
GLM-5.3Zhipu AI · 753B · Q4_KGGUF ↗ (External link)Original ↗ (External link)Too big~449.8 GB435.2 + 0.7 + 13.9
GLM-5.2Zhipu AI · 753B MoE · Q4_K_MGGUF ↗ (External link)Original ↗ (External link)Too big~448.3 GB433.8 + 0.7 + 13.9
DeepSeek-R1DeepSeek · 685B MoE · Q4_K_MGGUF ↗ (External link)Original ↗ (External link)Too big~389.4 GB376.7 + 0.5 + 12.2
Ornith 1.5 397BOrnith AI · 403B MoE · Q4_K_MGGUF ↗ (External link)Original ↗ (External link)Ollama ↗ (External link)Too big~235.4 GB227.5 + 0.2 + 7.7
GLM-5.3-FlashZhipu AI · 321B · Q4_KGGUF ↗ (External link)Original ↗ (External link)Too big~192.5 GB186.0 + 0.1 + 6.4
Hy3Tencent · 299B MoE · Q4_K_MGGUF ↗ (External link)Original ↗ (External link)Too big~174.6 GB166.3 + 2.5 + 5.8
DeepSeek-V4-FlashDeepSeek · 304B MoE · Q4_KGGUF ↗ (External link)Original ↗ (External link)Too big~150.0 GB144.4 + 0.4 + 5.2
Qwen3.8 Flash-NextAlibaba Qwen · 180B MoE · Q4_KGGUF ↗ (External link)Original ↗ (External link)Too big~107.8 GB103.7 + 0.2 + 4.0
Llama 3.1 8BMeta · 8B · Q4_K_MGGUF ↗ (External link)Original ↗ (External link)Ollama ↗ (External link)Unknown—
Llama 3.2 3BMeta · 3B · Q4_K_MGGUF ↗ (External link)Original ↗ (External link)Ollama ↗ (External link)Unknown—

All memory figures are estimates: measured quantized file size + computed context memory + runtime overhead, with a 12% safety margin on your hardware. Real usage varies with runtime version and settings.

GGUF is the quantized single-file format local runtimes load (Ollama, LM Studio, llama.cpp); our memory figures are measured from the linked GGUF file. The original repo holds full-precision safetensors — a much larger download meant for GPUs with far more memory.

Same memory, different speed

No other card in our catalogue carries this much memory, so nothing here is a like-for-like alternative.

Radar Card

RTX 5090 as a Radar Card

The same machine as a collectible spec card: a hologram drawn from its dimensions and parts list, sourced facts with their status, and a number issued in claim order that stays in your browser.

Open the card →

When this card is not enough

Own the middle.

DGX Spark and GB10-class systems put 128 GB of unified memory on your desk. Rational when large local models are your daily routine.

Read our DGX Spark profile →

Path C — Elastic compute

Rent the spike.

Occasional heavy job? Rent an H100 for the afternoon instead of buying hardware that idles the rest of the year.

RTX 5090 rental price per hour →

Keep reading