Computed with the open FitLLM engine — accurate per-layer KV-cache modeling, not a naive estimate. Updated 2026-10-23.
FitLLM compares fit — what loads in memory — computed from official config.json. These are a floor, not a guarantee; speed and power are not estimated.
| RTX 5090 | RTX 4090 | |
|---|---|---|
| VRAM | 32 GB | 24 GB |
| Memory bandwidth (speed, not estimated) | 1792 GB/s | 1008 GB/s |
| Model | RTX 5090 | RTX 4090 |
|---|---|---|
| Qwen 3.8 27B | ✅ rec. max 83K · 22.9/32 GB | ⚠️ rec. max 7K · 22.9/24 GB |
| Qwen 3.8 2.4T-A95B | ❌ won't fit · 1565/32 GB | ❌ won't fit · 1565/24 GB |
| Laguna XS 2.1 | ✅ rec. max 95K · 24.0/32 GB | ❌ limit 8K, no headroom · 24.0/24 GB |
| Laguna S 2.1 | ❌ won't fit · 103/32 GB | ❌ won't fit · 103/24 GB |
| Hy3 | ❌ won't fit · 196/32 GB | ❌ won't fit · 196/24 GB |
| GLM-5.2 | ❌ won't fit · 484/32 GB | ❌ won't fit · 484/24 GB |
| GLM-4.7-Flash | ✅ rec. max 96K · 22.6/32 GB | ✅ rec. max 10K · 22.6/24 GB |
| gpt-oss-20b | ✅ rec. max 128K · 14.6/32 GB | ✅ rec. max 128K · 14.6/24 GB |
| gpt-oss-120b | ❌ won't fit · 68.1/32 GB | ❌ won't fit · 68.1/24 GB |
| Qwen 3.6 35B-A3B | ✅ rec. max 98K · 25.7/32 GB | ❌ won't fit · 25.7/24 GB |
| Qwen 3.6 27B | ✅ rec. max 83K · 22.9/32 GB | ⚠️ rec. max 7K · 22.9/24 GB |
| Qwen-AgentWorld-35B-A3B | ✅ rec. max 119K · 24.6/32 GB | ❌ won't fit · 24.6/24 GB |
| Gemma 4 31b | ✅ rec. max 60K · 24.3/32 GB | ❌ limit 6K, no headroom · 24.3/24 GB |
| Gemma 4 26b A4B | ✅ rec. max 200K · 20.4/32 GB | ✅ rec. max 54K · 20.4/24 GB |
| Gemma 4 12b | ✅ rec. max 256K · 10.7/32 GB | ✅ rec. max 262K · 10.7/24 GB |
| Llama-3.1-8B-Instruct | ✅ rec. max 128K · 8.5/32 GB | ✅ rec. max 92K · 8.5/24 GB |
| Llama-3.2-3B-Instruct | ✅ rec. max 128K · 5.4/32 GB | ✅ rec. max 122K · 5.4/24 GB |
| MiniCPM5-1B | ✅ rec. max 128K · 3.2/32 GB | ✅ rec. max 128K · 3.2/24 GB |
Only the RTX 5090 runs: Qwen 3.8 27B, Laguna XS 2.1, Qwen 3.6 35B-A3B, Qwen 3.6 27B, Qwen-AgentWorld-35B-A3B, Gemma 4 31b.
For local LLMs, more VRAM means more models and longer context. Match the card to the model you actually want to run — see the per-model fit pages.
All numbers are computed by the open-source fitllm-engine (MIT) from official model config.json values — reproduce or audit them yourself. Estimates; real usage varies with runtime (llama.cpp / MLX / Ollama), driver and display. Found a mismatch? Report it. · FitLLM home