Every open model in our directory, ranked by how well it runs on MSI EdgeXpert (128GB) at 4-bit. Grades weigh whether the model fits in memory and how fast it should run.
MSI EdgeXpert has 128 GB of VRAM, enough to run 155 of the 169 open models we track at 4-bit. The largest that still fits well is Kimi K2 Instruct 0905 (1000B), and the top-rated is Qwen3-30B-A3B at about 41 tok/s.
155 of 169 models run comfortably on MSI EdgeXpert.
See full MSI EdgeXpert specs and pricingThe top models this device can run at 4-bit, ranked by fit and speed.
| Model | Grade | Speed | VRAM |
|---|---|---|---|
| Qwen3-30B-A3BAlibaba | SS | 40.8 tok/s | 5.4 GB |
| LFM2.5-8B-A1BLiquid AI | AA | 75.6 tok/s | 2.9 GB |
| Llama 3 8B InstructMeta | AA | 38.8 tok/s | 5.7 GB |
| Carnice-9b for Hermes agentkai-os | AA | 36.5 tok/s | 6.0 GB |
| PersonaPlex 7B |
| AA |
| 45.9 tok/s |
| 4.8 GB |
| Llama 2 7B ChatMeta | AA | 45.9 tok/s | 4.8 GB |
| Gemma 4 E2B ITGoogle | AA | 59.3 tok/s | 3.7 GB |
| VibeThinker-3BWeiboAI | AA | 57.6 tok/s | 3.8 GB |
| North Mini CodeCohere | AA | 26.2 tok/s | 8.4 GB |
| Nemotron 3 Nano OmniNVIDIA | AA | 25.8 tok/s | 8.5 GB |
| Qwen3.6 35B-A3BAlibaba | AA | 25.8 tok/s | 8.5 GB |
| Qwen3.5-35B-A3BAlibaba | AA | 25.8 tok/s | 8.5 GB |
Qwen3-30B-A3BAlibaba | 30B(3B active) | SS | 40.8 tok/s | 5.4 GB | |
LFM2.5-8B-A1BLiquid AI | 8.3B(1.5B active) | AA | 75.6 tok/s | 2.9 GB | |
| 8B | AA | 38.8 tok/s | 5.7 GB | ||
| 9B | AA | 36.5 tok/s | 6.0 GB | ||
PersonaPlex 7BNVIDIA | 7B | AA | 45.9 tok/s | 4.8 GB | |
Llama 2 7B ChatMeta | 7B | AA | 45.9 tok/s | 4.8 GB | |
Gemma 4 E2B ITGoogle | 2B | AA | 59.3 tok/s | 3.7 GB | |
VibeThinker-3BWeiboAI | 3B | AA | 57.6 tok/s | 3.8 GB | |
Ad CodeRabbitAI-powered Code Reviews. Cut review time & bugs in half, instantly.Try for Free | |||||
North Mini CodeCohere | 30B(3B active) | AA | 26.2 tok/s | 8.4 GB | |
Nemotron 3 Nano OmniNVIDIA | 30B(3B active) | AA | 25.8 tok/s | 8.5 GB | |
Qwen3.6 35B-A3BAlibaba | 35B(3B active) | AA | 25.8 tok/s | 8.5 GB | |
Qwen3.5-35B-A3BAlibaba | 35B(3B active) | AA | 25.8 tok/s | 8.5 GB | |
Mistral 7B InstructMistral AI | 7B | AA | 34.4 tok/s | 6.4 GB | |
Llama 2 13B ChatMeta | 13B | AA | 26.0 tok/s | 8.5 GB | |
Gemma 4 E4B ITGoogle | 4B | AA | 31.8 tok/s | 6.9 GB | |
Gemma 3 4B ITGoogle | 4B | AA | 31.8 tok/s | 6.9 GB | |
Ad Vast.aiAffordable on-demand GPU rentals for training and inference. Pick from thousands of hosts.Rent a GPU | |||||
Mixtral 8x7B InstructMistral AI | 46.7B(12.9B active) | BB | 19.3 tok/s | 11.4 GB | |
DiffusionGemma 26B-A4BGoogle | 25.2B(3.8B active) | BB | 21.0 tok/s | 10.5 GB | |
Gemma 4 26B-A4B ITGoogle | 26B(4B active) | BB | 20.0 tok/s | 11.0 GB | |
Mistral Large 3 675BMistral AI | 675B(41B active) | BB | 3.3 tok/s | 66.3 GB | |
GLM-4.6Z.ai | 355B(32B active) | BB | 3.1 tok/s | 70.3 GB | |
DeepSeek-V3DeepSeek | 671B(37B active) | BB | 3.7 tok/s | 59.8 GB | |
DeepSeek-R1DeepSeek | 671B(37B active) | BB | 3.7 tok/s | 59.8 GB | |
DeepSeek-V3.1DeepSeek | 671B(37B active) | BB | 3.7 tok/s | 59.8 GB | |
Ad RunPodServerless and dedicated GPU cloud built for AI workloads. Spin up instances in seconds.Launch on RunPod | |||||
DeepSeek-V3.2DeepSeek | 685B(37B active) | BB | 3.7 tok/s | 59.8 GB | |
Model Compatibility
Use the calculator to weigh this device against any other, model by model, with speed and fit scores.
Straight answers to the questions we hear most often.
Can't find what you're looking for? Send us a message
Qwen3-30B-A3B is the highest-rated model that runs on MSI EdgeXpert at 4-bit, at about 41 tok/s. In total, MSI EdgeXpert can run 155 of the 169 open models we track.
We size each model at 4-bit (Q4_K_M), add the KV cache for its context length and a fixed runtime overhead, then compare that to this device VRAM. Each model is graded on how comfortably it fits and how fast it should run on this card.
It is an estimate of decode speed in tokens per second at 4-bit, based on this device memory bandwidth and each model size. Real speed depends on your runtime, batch size, and context length, so use it to compare models, not as a hard promise.
Lower-bit quantization shrinks a model so more of them fit. These grades already assume 4-bit, which is the common local default. Going below 4-bit can squeeze a larger model in at some quality cost; going above needs more VRAM and may not fit.
Mixture-of-experts models only activate a fraction of their parameters at a time, so their runtime memory is closer to the active size than the headline size. We score fit on the active parameters, which is why some very large models still fit.
