will it local
GeForce flagship: GB202, 21760 CUDA, 512-bit GDDR7 at 28 Gbps = 1792 GB/s (TechPowerUp, live-pinned 0905). The single fastest GeForce for local inference — 32GB at ~1.8 TB/s covers 32B Q4 with real context. Buy it for the bandwidth + VRAM combo, not the tensor-core marketing (llama.cpp GGUF decode is bandwidth-bound).
computed band 190.0-280.0 tok/s for 8B Q4_K_M (roofline, 1792.0 GB/s VRAM); community: 220 tok/s Meta Llama 3.1 8B Instruct Q4_K_M (https://www.kunalganglani.com/llm-benchmarks); community: 66.3 tok/s Meta Llama 3.1 8B Instruct Q4_K_M (https://www.localscore.ai/accelerator/155)
Effective decode window: 0.55–0.80 of 1792 GB/s nominal → ~986–1434 GB/s effective (llama.cpp decode, Q4_K_M basis; card-specific window from cited community benches)
Community check — localmaxxing.com medians (6 buckets): localmaxxing.com
→ ~220 tok/s · 3-4B Q8 · 3 runs · just below our band (0% under the low edge)
→ ~88 tok/s · 12-15B Q8 · 5 runs · inside our computed band
→ ~73 tok/s · 30-33B Q3 · 9 runs · just above our band (5% over the high edge)
→ ~67 tok/s · 30-33B Q4 · 13 runs · inside our computed band
→ ~60 tok/s · 30-33B Q5 · 7 runs · just above our band (2% over the high edge)
→ ~51 tok/s · 30-33B Q6 · 6 runs · inside our computed band
→ 13b-q4 full 14.6 GB needed of 32 GB usable — headroom for context
→ 3b-q4 full 2.8 GB needed of 32 GB usable — headroom for context
→ 4b-q4 full 3.7 GB needed of 32 GB usable — headroom for context
→ 8b-q4 full 6.1 GB needed of 32 GB usable — headroom for context
→ 220 tok/s · Meta Llama 3.1 8B Instruct Q4_K_M Kunal Ganglani bench DB (measured, llama.cpp) · source
→ 66.3 tok/s · Meta Llama 3.1 8B Instruct Q4_K_M LocalScore median (early sample, skewed low) · source
Nominal bandwidth: 1792.0 GB/s — see the effective decode window above
→ qwen3:0.6b on Raspberry Pi 5 + AI Kit (Hailo-8L) ($110) 0.6B · 0.6GB · needs ~1GB (4k ctx)
→ qwen3:1.7b on Raspberry Pi 5 + AI Kit (Hailo-8L) ($110) 1.7B · 1.4GB · needs ~1.8GB (4k ctx)
→ qwen3:4b on Raspberry Pi 5 + AI Kit (Hailo-8L) ($110) 4B · 2.6GB · needs ~3.2GB (4k ctx)
→ llama3.2:1b on Raspberry Pi 5 + AI Kit (Hailo-8L) ($110) 1B · 1.3GB · needs ~1.4GB (4k ctx)
→ llama3.2:3b on Raspberry Pi 5 + AI Kit (Hailo-8L) ($110) 3B · 2.0GB · needs ~2.4GB (4k ctx)
→ gemma3:1b on Raspberry Pi 5 + AI Kit (Hailo-8L) ($110) 1B · 1.0GB · needs ~1.1GB (4k ctx)
→ phi4:mini on Raspberry Pi 5 + AI Kit (Hailo-8L) ($110) mini 3.8B · 2.5GB · needs ~3GB (4k ctx)
→ moondream:2b on Raspberry Pi 5 + AI Kit (Hailo-8L) ($110) 2B · 1.7GB · needs ~3.7GB (4k ctx)
→ whisper:tiny on Raspberry Pi 5 + AI Kit (Hailo-8L) ($110) tiny · 0.1GB · needs ~2.1GB (4k ctx)
→ whisper on Raspberry Pi 5 + AI Kit (Hailo-8L) ($110) base · 0.3GB · needs ~2.3GB (4k ctx)
→ whisper:small on Raspberry Pi 5 + AI Kit (Hailo-8L) ($110) small · 0.9GB · needs ~2.9GB (4k ctx)
→ whisper:medium on Raspberry Pi 5 + AI Kit (Hailo-8L) ($110) medium · 3.1GB · needs ~5.1GB (4k ctx)
→ kokoro:82m on Raspberry Pi 5 + AI Kit (Hailo-8L) ($110) 82M · 0.3GB · needs ~2.3GB (4k ctx)
→ gemma3:4b on Raspberry Pi 5 + AI Kit (Hailo-8L) ($110) 4B · 3.3GB · needs ~3.8GB (4k ctx)
→ deepseek-r1:7b on Raspberry Pi 5 + AI Kit (Hailo-8L) ($110) distill 7B · 4.7GB · needs ~4.9GB (4k ctx)
→ mistral:7b on Raspberry Pi 5 + AI Kit (Hailo-8L) ($110) 7B · 4.1GB · needs ~4.6GB (4k ctx)
→ llava:7b on Raspberry Pi 5 + AI Kit (Hailo-8L) ($110) 7B · 4.7GB · needs ~6.7GB (4k ctx)
→ qwen3:8b on Tesla P100 (16GB, used) ($135) 8B · 5.2GB · needs ~5.8GB (4k ctx)
→ qwen2.5vl:7b on Tesla P100 (16GB, used) ($135) 7B · 5.6GB · needs ~5.8GB (4k ctx)
→ llama3.1:8b on Tesla P100 (16GB, used) ($135) 8B · 4.9GB · needs ~5.4GB (4k ctx)
→ deepseek-r1:8b on Tesla P100 (16GB, used) ($135) distill 8B · 4.9GB · needs ~5.4GB (4k ctx)
→ whisper:large-v3 on Tesla P100 (16GB, used) ($135) large-v3 · 6.2GB · needs ~8.2GB (4k ctx)
→ mistral-nemo:12b on Tesla P100 (16GB, used) ($135) 12B · 7.1GB · needs ~7.7GB (4k ctx)
→ sdxl on Tesla P100 (16GB, used) ($135) SDXL base · 6.9GB · needs ~8.9GB (4k ctx)
→ llava:13b on Tesla P100 (16GB, used) ($135) 13B · 8.0GB · needs ~11.1GB (4k ctx)
→ gemma3:12b on Tesla P100 (16GB, used) ($135) 12B · 8.1GB · needs ~9.6GB (4k ctx)
→ qwen3:14b on Tesla P100 (16GB, used) ($135) 14B · 9.3GB · needs ~9.9GB (4k ctx)
→ deepseek-r1:14b on Tesla P100 (16GB, used) ($135) distill 14B · 9.0GB · needs ~9.8GB (4k ctx)
→ phi4:14b on Tesla P100 (16GB, used) ($135) 14B · 9.1GB · needs ~9.9GB (4k ctx)
→ gpt-oss:20b on Tesla P40 (24GB, used) ($290) 20B MoE · 14GB · needs ~14.2GB (4k ctx)
→ devstral:24b on Tesla P40 (24GB, used) ($290) 24B · 14.6GB · needs ~15.2GB (4k ctx)
→ qwen3.8:27b on Mac mini M4 (32GB) ($999) 27B · 16.5GB · needs ~18.5GB (4k ctx)
→ gemma3:27b on Mac mini M4 (32GB) ($999) 27B · 17GB · needs ~18.9GB (4k ctx)
→ flux:schnell on Mac mini M4 (32GB) ($999) schnell fp8 · 17GB · needs ~19GB (4k ctx)
→ qwen3:30b-a3b on Mac mini M4 (32GB) ($999) 30B-A3B MoE · 18.6GB · needs ~19GB (4k ctx)
→ qwen3-coder:30b-a3b on Mac mini M4 (32GB) ($999) 30B-A3B MoE · 18.6GB · needs ~19GB (4k ctx)
→ qwen3:32b on Mac mini M4 (32GB) ($999) 32B · 20GB · needs ~21GB (4k ctx)
→ deepseek-r1:32b on Mac mini M4 (32GB) ($999) distill 32B · 20GB · needs ~21GB (4k ctx)
→ qwen2.5vl:32b on Mac mini M5 Pro (64GB) ($1669) 32B · 22GB · needs ~23GB (4k ctx)
→ flux:dev on Mac mini M5 Pro (64GB) ($1669) dev fp8 · 23GB · needs ~25GB (4k ctx)
→ mixtral:8x7b on Mac mini M5 Pro (64GB) ($1669) 8x7B MoE · 26GB · needs ~26.5GB (4k ctx)
—
⚠ 575W TDP on a 16-pin connector — you need a 950W-class PSU and a case that takes a 304mm triple-slot-flow card; the burned-connector panic of early 2025 means use the native 12V-2x6 cable, not adapters
⚠ 32GB does NOT fit 70B Q4 — the flagship VRAM story is 32B-class Q4 with big context, or 70B at heavy offload (not worth it on one card)
⚠ MSRP $1,999 was never the street price; 2026 RAMmageddon keeps 5090s far above MSRP — a used 2× 3090 rig often beats it on tok/s-per-dollar for 30B+
⚠ LocalScore median (66.3 tok/s 8B) reads far below community rows (213-220) — early single-sample medians on new silicon skew low; band is to contain both
⚠ No NVLink on Blackwell — multi-5090 runs over PCIe (pipeline-parallel fine, tensor-parallel slower than 3090 NVLink pairs)
$16499 — 96GB Blackwell — 70B Q4 single-card territory at workstation pricing
$50000 — 4× PRO 6000 if you actually need 70B+ dense at speed