GMKtec EVO-X2 (128GB unified)
The $/GB winner: runs 70B-class and quantized big-MoE (V4 Flash tier) that no consumer GPU fits
Listed H200 NVL rates (141 GB VRAM) from tracked provider APIs, sorted per GPU-hour. Source configurations may contain several GPUs or different editions. Availability is not checked; confirm the configuration and total charge with the provider.
Quotes refreshed 2026-11-22| Provider | Tier | $/hr per GPU | Source configuration | Instance charge | Availability | |
|---|---|---|---|---|---|---|
| RunPod | community | $0.50 | Source GPU SKU · NVIDIA H200 NVL | Instance charge not supplied | Availability not checked | Check provider |
| RunPod | secure | $3.79 | Source GPU SKU · NVIDIA H200 NVL | Instance charge not supplied | Availability not checked | Check provider |
Models from our tracked open-weight set that fit in 141 GB, with KV-cache headroom included. Bigger models need multi-GPU nodes (math in the calculator).
| Model | Params | Fits at | VRAM needed | Cheapest API $/1M out |
|---|---|---|---|---|
| Qwen3.8 27B | 27.8B | FP8 | 39 GB | compare → |
| GLM-4.7 Flash | 31B | FP8 | 43 GB | compare → |
| Qwen3.6 35B A3B | 36B | FP8 | 50 GB | compare → |
| Qwen3.6 27B | 27.8B | FP8 | 39 GB | compare → |
| Qwen3.5 35B A3B | 36B | FP8 | 50 GB | compare → |
| Qwen3.5 27B | 27.8B | FP8 | 39 GB | compare → |
| Qwen3 32B | 32.8B | FP8 | 46 GB | compare → |
| Qwen3 30B A3B | 30.5B | FP8 | 42 GB | compare → |
| Mistral Small 3.1 | 24B | FP8 | 33 GB | |
| Nemotron 3 Nano | 30B | FP8 | 42 GB | compare → |
| IBM Granite 4.1 30B | 30B | FP8 | 42 GB | |
| IBM Granite 4.1 8B | 8.8B | FP8 | 13 GB | compare → |
| GPT-OSS 20B | 21B | FP8 | 29 GB | compare → |
| Llama 3.3 70B | 70B | FP8 | 97 GB | compare → |
| Llama 3.1 8B | 8B | FP8 | 11 GB | compare → |
| Llama 3.2 3B | 3.2B | FP8 | 5 GB | compare → |
| Gemma 3 27B | 27B | FP8 | 38 GB | compare → |
| Gemma 3 12B | 12B | FP8 | 17 GB | compare → |
| Phi-4 | 14.7B | FP8 | 21 GB | compare → |
| Mixtral 8x7B | 46.7B | FP8 | 65 GB | compare → |
| DeepSeek R1 Distill Llama 70B | 70B | FP8 | 97 GB | compare → |
| DeepSeek R1 Distill Qwen 32B | 32.8B | FP8 | 46 GB | compare → |
| GLM-4.5 Air | 106B | INT4 | 73 GB | compare → |
| Mistral Medium 3.5 | 128B | INT4 | 88 GB | compare → |
| Mistral Small 4 | 119B | INT4 | 82 GB | |
| Leanstral 1.5 | 119B | INT4 | 82 GB | |
| Nemotron 3 Super | 120B | INT4 | 83 GB | compare → |
| GPT-OSS 120B | 117B | INT4 | 81 GB | compare → |
| Llama 4 Scout | 109B | INT4 | 75 GB | compare → |
Compare local hardware or estimate rental costs in the calculator.
The $/GB winner: runs 70B-class and quantized big-MoE (V4 Flash tier) that no consumer GPU fits
Bandwidth-per-dollar king for the A3B-MoE and 27-32B dense workhorses
The startup standard: only sub-$10K single card that runs 70B+ comfortably
As an Amazon Associate we earn from qualifying purchases · hardware links may be referral links · picks are editorial, not paid · disclosure
The lowest listed rate in our tracked snapshot is $0.50/GPU-hour (community, RunPod), observed 2026-11-22. Source configuration: Source GPU SKU · NVIDIA H200 NVL. Instance charge not supplied. Availability not checked. This is a normalized comparison, not a provider-wide minimum or a confirmed bookable offer.
With 141 GB of VRAM, a single H200 NVL fits models up to roughly 100B parameters at FP8 or 201B at INT4 quantization, including KV-cache headroom.
Lowest tracked listed H200 NVL rate: $0.50/GPU-hour, RunPod (community), Source GPU SKU · NVIDIA H200 NVL. Instance charge not supplied. Availability not checked. Observed 2026-11-22 (llmhosting.ai). Provider data terms apply; a link back does not grant redistribution rights.