HARDWARE / VRAM
Datacenter · 80 GB
Large MoE models and high-throughput servers, still bounded by quantization and cache.
H100 / H200 / MI300X class80 GB
01 / PRACTICAL CEILING
Llama 3.1 70B
The highest listed result may be a tight or offloaded fit. For daily use, prefer the first result marked Comfortable and keep context modest.
02 / DEFAULT RUNTIME
TensorRT-LLM
Maximum NVIDIA throughput after engine tuning. NVIDIA-only and sensitive to supported GPU, CUDA, model architecture and engine-build settings.
Apri guida03 / MODEL ENVELOPE
Modelli open-weight rappresentativi
Meta
Llama 3.1 70B
70BGGUF Q4
- Memoria stimata
- ~46 GB
- Contesto
- 128K
Alibaba Qwen
Qwen3 32B
32BGGUF Q4
- Memoria stimata
- ~24.5 GB
- Contesto
- 32K+
Alibaba Qwen
Qwen3 30B-A3B
30B3B active · GGUF Q4 MoE
- Memoria stimata
- ~22.5 GB
- Contesto
- 32K+
Google
Gemma 3 27B
27BINT4
- Memoria stimata
- ~20.5 GB
- Contesto
- 128K
Mistral AI
Devstral Small 2 24B
24BGGUF Q4 / BF16
- Memoria stimata
- ~18.5 GB
- Contesto
- 256K
OpenAI
gpt-oss-20b
21B3.6B active · MXFP4
- Memoria stimata
- ~16 GB
- Contesto
- 128K
Attenzione
Fit is based on estimated total model memory. Driver support, KV cache, multimodal projectors, concurrency, and desktop applications can all reduce available headroom. Current fit: Llama 3.1 70B (Comfortable), Qwen3 32B (Comfortable), Qwen3 30B-A3B (Comfortable), Gemma 3 27B (Comfortable), Devstral Small 2 24B (Comfortable), gpt-oss-20b (Comfortable).