Every open model in our directory, ranked by how well it runs on Intel Gaudi 2 AI Accelerator (96GB) at 4-bit. Grades weigh whether the model fits in memory and how fast it should run.
Intel Gaudi 2 AI Accelerator has 96 GB of VRAM, enough to run 145 of the 169 open models we track at 4-bit. The largest that still fits well is Kimi K2 Instruct (1000B), and the top-rated is Qwen3.5-397B-A17B at about 43 tok/s.
145 of 169 models run comfortably on Intel Gaudi 2 AI Accelerator.
See full Intel Gaudi 2 AI Accelerator specs and pricingThe top models this device can run at 4-bit, ranked by fit and speed.
| Model | Grade | Speed | VRAM |
|---|---|---|---|
| Qwen3.5-397B-A17BAlibaba | SS | 42.9 tok/s | 46.0 GB |
| Llama 3 70B InstructMeta | SS | 43.2 tok/s | 45.7 GB |
| Qwen 3.5 OmniAlibaba | SS | 43.7 tok/s | 45.2 GB |
| Mixtral 8x22B InstructMistral AI | SS | 45.3 tok/s | 43.6 GB |
| Llama 2 70B ChatMeta | SS | 45.5 tok/s | 43.4 GB |
| GLM-4.5Z.ai | SS | 38.1 tok/s | 51.8 GB |
| Kimi K2 InstructMoonshot AI | SS | 38.1 tok/s | 51.8 GB |
| GLM-4.7Z.ai | SS | 37.5 tok/s | 52.6 GB |
| Gemma 3 27B ITGoogle | SS | 45.0 tok/s | 43.8 GB |
| Qwen3-235B-A22BAlibaba | SS | 54.3 tok/s | 36.3 GB |
| Mistral Small 3 24BMistral AI | SS | 50.6 tok/s | 39.0 GB |
| Hy3Tencent | SS | 35.4 tok/s | 55.7 GB |
Qwen3.5-397B-A17BAlibaba | 397B(17B active) | SS | 42.9 tok/s | 46.0 GB | |
| 70B | SS | 43.2 tok/s | 45.7 GB | ||
Qwen 3.5 OmniAlibaba | 397B(17B active) | SS | 43.7 tok/s | 45.2 GB | |
Mixtral 8x22B InstructMistral AI | 141B(39B active) | SS | 45.3 tok/s | 43.6 GB | |
Llama 2 70B ChatMeta | 70B | SS | 45.5 tok/s | 43.4 GB | |
GLM-4.5Z.ai | 355B(32B active) | SS | 38.1 tok/s | 51.8 GB | |
Kimi K2 InstructMoonshot AI | 1000B(32B active) | SS | 38.1 tok/s | 51.8 GB | |
GLM-4.7Z.ai | 358B(32B active) | SS | 37.5 tok/s | 52.6 GB | |
Ad CodeRabbitAI-powered Code Reviews. Cut review time & bugs in half, instantly.Try for Free | |||||
Gemma 3 27B ITGoogle | 27B | SS | 45.0 tok/s | 43.8 GB | |
Qwen3-235B-A22BAlibaba | 235B(22B active) | SS | 54.3 tok/s | 36.3 GB | |
Mistral Small 3 24BMistral AI | 24B | SS | 50.6 tok/s | 39.0 GB | |
Hy3Tencent | 295B(21B active) | SS | 35.4 tok/s | 55.7 GB | |
Qwen3-32BAlibaba | 32.8B | SS | 36.6 tok/s | 53.9 GB | |
LLaMA 65BMeta | 65B | SS | 50.2 tok/s | 39.3 GB | |
Qwen3.5-122B-A10BAlibaba | 122B(10B active) | SS | 72.3 tok/s | 27.3 GB | |
DeepSeek-V3DeepSeek | 671B(37B active) | SS | 33.0 tok/s | 59.8 GB | |
Ad PayPerQPay-per-query access to top LLMs without a subscription. Use any model on demand.Try PayPerQ | |||||
DeepSeek-R1DeepSeek | 671B(37B active) | SS | 33.0 tok/s | 59.8 GB | |
DeepSeek-V3.1DeepSeek | 671B(37B active) | SS | 33.0 tok/s | 59.8 GB | |
DeepSeek-V3.2DeepSeek | 685B(37B active) | SS | 33.0 tok/s | 59.8 GB | |
minimax-m2.5MiniMax | 230B(10B active) | SS | 86.9 tok/s | 22.7 GB | |
Gemma 4 12BGoogle | 12B | SS | 61.6 tok/s | 32.0 GB | |
Gemma 4 12B Coderyuxinlu1 | 12B | SS | 61.6 tok/s | 32.0 GB | |
Mistral Large 3 675BMistral AI | 675B(41B active) | SS | 29.8 tok/s | 66.3 GB | |
Mixtral 8x7B InstructMistral AI | 46.7B(12.9B active) | SS | 173.6 tok/s | 11.4 GB | |
Ad Vast.aiAffordable on-demand GPU rentals for training and inference. Pick from thousands of hosts.Rent a GPU | |||||
Falcon 40B InstructTechnology Innovation Institute | 40B | SS | 81.0 tok/s | 24.4 GB | |
Model Compatibility
Use the calculator to weigh this device against any other, model by model, with speed and fit scores.
Qwen3.5-397B-A17B is the highest-rated model that runs on Intel Gaudi 2 AI Accelerator at 4-bit, at about 43 tok/s. In total, Intel Gaudi 2 AI Accelerator can run 145 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.
