will it local
Pre-RDNA Polaris (gfx803): no ROCm support on modern stacks — llama.cpp Vulkan (RADV) is the only viable GPU route 4GB VRAM is the hard ceiling: 7B Q4_K_M (~4.1GB) does NOT fit fully on-GPU; realistic lane is 3-4B models at Q4 or 7B at IQ3/IQ4 with partial offload Localmaxxing numbers are a mixed-model aggregate (224 runs, includes 2x multi-GPU); no clean 7B Q4 single-card median published — treat 44 tok/s as small-model territory, 7B class will be roughly RX 580-class (~8-15 tok/s) only with partial offload Many used cards are ex-mining; check memory temps and fan wear Some 4GB cards are BIOS-flashable from/to 8GB variants; verify actual VRAM with vulkaninfo
computed band 3.5-9.2 tok/s for 8B Q4_K_M (roofline, 224.0 GB/s); community: 44.0 median (best 125, min 3.8) tok/s Mixed small models (community speed tests, model/quant varies per run) mixed (https://www.localmaxxing.com/en/hardware/DISCRETE_GPU%3Arx%20570%204gb?name=RX%20570%204GB&gpuCount=1)
Effective decode window: 0.08–0.21 of 224 GB/s nominal → ~18–47 GB/s effective (llama.cpp decode, Q4_K_M basis; arch window, community-calibrated)
Community check — localmaxxing.com medians (1 bucket): localmaxxing.com
→ ~28 tok/s · 3-4B Q4 · 3 runs · above our band — small sample; documented outlier, computed window stands
→ 3b-q4 full 2.8 GB needed of 4 GB usable — headroom for context
→ 4b-q4 tight 3.7 of 4 GB usable — keep context modest
→ 8b-q4 no 6.1 GB vs 4 GB usable — does not fit
→ 44.0 median (best 125, min 3.8) tok/s · Mixed small models (community speed tests, model/quant varies per run) mixed · source
Nominal bandwidth: 224.0 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)
→ 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 distill 7B — 4.9 GB vs 4 GB usable — partial CPU offload, expect large speed loss
→ mistral:7b 7B — 4.6 GB vs 4 GB usable — partial CPU offload, expect large speed loss
⚠ Pre-RDNA Polaris (gfx803): no ROCm support on modern stacks — llama.cpp Vulkan (RADV) is the only viable GPU route
⚠ 4GB VRAM is the hard ceiling: 7B Q4_K_M (~4.1GB) does NOT fit fully on-GPU; realistic lane is 3-4B models at Q4 or 7B at IQ3/IQ4 with partial offload
⚠ Localmaxxing numbers are a mixed-model aggregate (224 runs, includes 2x multi-GPU); no clean 7B Q4 single-card median published — treat 44 tok/s as small-model territory, 7B class will be roughly RX 580-class (~8-15 tok/s) only with partial offload
⚠ Many used cards are ex-mining; check memory temps and fan wear
⚠ Some 4GB cards are BIOS-flashable from/to 8GB variants; verify actual VRAM with vulkaninfo