will it local

Best local LLMs for the GeForce MX350

⚠ Laptop part — mobile clocks and thermals run below desktop sibling speeds; treat the band as optimistic for sustained load.

Speed

computed band 3.1-5.4 tok/s for 8B Q4_K_M (roofline, 56 GB/s VRAM)

Effective decode window: 0.28–0.50 of 56 GB/s nominal → ~16–28 GB/s effective (llama.cpp decode, Q4_K_M basis; arch window, community-calibrated)

What fits (computed)

→ 13b-q4 no 14.6 GB vs 2 GB usable — does not fit

→ 3b-q4 no 2.8 GB vs 2 GB usable — does not fit

→ 4b-q4 no 3.7 GB vs 2 GB usable — does not fit

→ 8b-q4 no 6.1 GB vs 2 GB usable — does not fit

Cited community benches

no single-card bench published — band is computed (see notes)

Runs fully in memory (machine alternatives)

Nominal bandwidth: 56 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)

→ llama3.2:1b on Raspberry Pi 5 + AI Kit (Hailo-8L) ($110) 1B · 1.3GB · needs ~1.4GB (4k ctx)

→ gemma3:1b on Raspberry Pi 5 + AI Kit (Hailo-8L) ($110) 1B · 1.0GB · needs ~1.1GB (4k ctx)

Borderline — runs, but offloads

→ llama3.2:3b 3B — 2.4 GB vs 2 GB usable — partial CPU offload, expect large speed loss

→ whisper:tiny tiny — 2.1 of 2 GB usable — barely over; real with q8 KV cache (halves KV) or shorter context

→ whisper base — 2.3 GB vs 2 GB usable — partial CPU offload, expect large speed loss

→ kokoro:82m 82M — 2.3 GB vs 2 GB usable — partial CPU offload, expect large speed loss

Gotchas

⚠ 2 GB VRAM hard wall; 64-bit bus at 7 Gbps = 56 GB/s (TPU), barely ahead of the MX150

⚠ GP107 Pascal: 640 CUDA cores, no tensor cores; pre-Turing CUDA deprecated by many current LLM stacks

⚠ Alias MX330 is a GP108 rehash: 384 cores and only 48 GB/s — slower than the MX350 it's grouped with

⚠ No reliable per-card tok/s benches found; not a usable local-LLM GPU

Where to go next

—