Olmo 3 · released · 7.43B parameters
24 Gated DeltaNet layers (30 heads, k 96 / v 192; no KV cache) + 8 full MHA layers (every 4th). Release month not stated (repos created Jan-Feb 2026). DeltaNet state assumed BF16.
Source configMemory for one conversation's context, on top of the weights.
| Context | FP16 | FP8 |
|---|---|---|
| 4,096 tokens | 0.50 GB | 0.26 GB |
| 8,192 tokens | 0.96 GB | 0.50 GB |
| 32,768 tokens | 3.78 GB | 1.90 GB |
| 65,536 tokens | 7.53 GB | 3.78 GB |
| Format | Bits | Weights |
|---|---|---|
| BF16 / FP16 | 16 | 13.8 GB |
| FP8 (W8A8) | 8 | 8.36 GB |
| INT8 (W8A8 / W8A16) | 8 | 8.36 GB |
| NVFP4 | 4.5 | 5.96 GB |
| INT4 AWQ / GPTQ | 4.25 | 5.79 GB |
| GGUF Q8_0 | 8.5 | 7.35 GB |
| GGUF Q6_K | 6.56 | 5.67 GB |
| GGUF Q5_K_M | 5.69 | 4.92 GB |
| GGUF Q4_K_M | 4.89 | 4.23 GB |
| GGUF Q3_K_M | 3.91 | 3.38 GB |
| GGUF Q2_K | 3.35 | 2.90 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× · 299 tok/s | 1× · 141 tok/s |
| H200 SXM 141GBFP8 (W8A8) | 1× · 209 tok/s | 1× · 90.7 tok/s |
| H100 SXM 80GBFP8 (W8A8) | 1× · 158 tok/s | 1× · 65.4 tok/s |
| Instinct MI300X 192GBFP8 (W8A8) | 1× · 225 tok/s | 1× · 99.0 tok/s |
| A100 SXM 80GBINT4 AWQ / GPTQ | 1× · 139 tok/s | 1× · 45.7 tok/s |
| L40S 48GBFP8 (W8A8) | 1× · 47.2 tok/s | 1× · 17.9 tok/s |
| RTX PRO 6000 Blackwell Server Edition 96GBFP8 (W8A8) | 1× · 83.3 tok/s | 1× · 32.5 tok/s |
| GeForce RTX 5090 32GBGGUF Q4_K_M | 1× · 123 tok/s | 1× · 36.2 tok/s |
| GeForce RTX 4090 24GBGGUF Q4_K_M | 1× · 79.4 tok/s | 2× · 36.8 tok/s |
| Apple M4 Max (40-core GPU, up to 128GB)GGUF Q4_K_M | 1× · 55.4 tok/s | 1× · 14.0 tok/s |
| DGX Spark (GB10, 128GB unified)FP8 (W8A8) | 1× · 15.5 tok/s | 1× · 5.7 tok/s |