Yi · released 2024-05 · 34.4B parameters
4K context; Yi-1.5-34B-Chat-16K variant has 16K.
Source configMemory for one conversation's context, on top of the weights.
| Context | FP16 | FP8 |
|---|---|---|
| 4,096 tokens | 0.94 GB | 0.47 GB |
| Format | Bits | Weights |
|---|---|---|
| BF16 / FP16 | 16 | 64.1 GB |
| FP8 (W8A8) | 8 | 33.7 GB |
| INT8 (W8A8 / W8A16) | 8 | 33.7 GB |
| NVFP4 | 4.5 | 20.5 GB |
| INT4 AWQ / GPTQ | 4.25 | 19.5 GB |
| GGUF Q8_0 | 8.5 | 34.0 GB |
| GGUF Q6_K | 6.56 | 26.3 GB |
| GGUF Q5_K_M | 5.69 | 22.8 GB |
| GGUF Q4_K_M | 4.89 | 19.6 GB |
| GGUF Q3_K_M | 3.91 | 15.7 GB |
| GGUF Q2_K | 3.35 | 13.4 GB |
One server, 8K context, FP8 on data center GPUs and GGUF Q4_K_M on consumer and Apple hardware.
| GPU | 1 user | 32 users |
|---|---|---|
| B200 SXM (HGX B200)FP8 (W8A8) | 1× · 100 tok/s | 1× · 58.6 tok/s |
| H200 SXM 141GBFP8 (W8A8) | 1× · 64.0 tok/s | 1× · 36.4 tok/s |
| H100 SXM 80GBFP8 (W8A8) | 1× · 46.0 tok/s | 1× · 25.9 tok/s |
| Instinct MI300X 192GBFP8 (W8A8) | 1× · 70.0 tok/s | 1× · 40.0 tok/s |
| A100 SXM 80GBINT4 AWQ / GPTQ | 1× · 47.4 tok/s | 1× · 20.4 tok/s |
| L40S 48GBFP8 (W8A8) | 1× · 12.5 tok/s | 2× · 12.6 tok/s |
| RTX PRO 6000 Blackwell Server Edition 96GBFP8 (W8A8) | 1× · 22.8 tok/s | 1× · 12.6 tok/s |
| GeForce RTX 5090 32GBGGUF Q4_K_M | 1× · 35.8 tok/s | 2× · 26.1 tok/s |
| GeForce RTX 4090 24GBGGUF Q4_K_M | 2× · 35.5 tok/s | 4× · 23.0 tok/s |
| Apple M4 Max (40-core GPU, up to 128GB)GGUF Q4_K_M | 1× · 14.1 tok/s | 1× · 2.8 tok/s |
| DGX Spark (GB10, 128GB unified)FP8 (W8A8) | 1× · 4.0 tok/s | 1× · 2.2 tok/s |