Qwen3-Next · released 2025-09 · 81.3B parameters (3.87B active per token)
Hybrid: 36 Gated DeltaNet layers (no KV cache) + 12 gated full-attention layers (every 4th). params include 1 MTP layer. Recurrent state assumed BF16 (config sets no mamba_ssm_dtype).
Source configMemory for one conversation's context, on top of the weights.
| Context | FP16 | FP8 |
|---|---|---|
| 4,096 tokens | 0.13 GB | 0.08 GB |
| 8,192 tokens | 0.22 GB | 0.13 GB |
| 32,768 tokens | 0.79 GB | 0.41 GB |
| 131,072 tokens | 3.04 GB | 1.54 GB |
| 262,144 tokens | 6.04 GB | 3.04 GB |
| Format | Bits | Weights |
|---|---|---|
| BF16 / FP16 | 16 | 151.5 GB |
| FP8 (W8A8) | 8 | 76.9 GB |
| INT8 (W8A8 / W8A16) | 8 | 76.9 GB |
| NVFP4 | 4.5 | 44.3 GB |
| INT4 AWQ / GPTQ | 4.25 | 41.9 GB |
| GGUF Q8_0 | 8.5 | 80.5 GB |
| GGUF Q6_K | 6.56 | 62.1 GB |
| GGUF Q5_K_M | 5.69 | 53.9 GB |
| GGUF Q4_K_M | 4.89 | 46.3 GB |
| GGUF Q3_K_M | 3.91 | 37.0 GB |
| GGUF Q2_K | 3.35 | 31.7 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× · 373 tok/s | 1× · 85.7 tok/s |
| H200 SXM 141GBFP8 (W8A8) | 1× · 281 tok/s | 1× · 54.0 tok/s |
| H100 SXM 80GBFP8 (W8A8) | 2× · 257 tok/s | 2× · 68.2 tok/s |
| Instinct MI300X 192GBFP8 (W8A8) | 1× · 299 tok/s | 1× · 59.1 tok/s |
| A100 SXM 80GBINT4 AWQ / GPTQ | 1× · 236 tok/s | 1× · 39.6 tok/s |
| L40S 48GBFP8 (W8A8) | 2× · 103 tok/s | 4× · 28.8 tok/s |
| RTX PRO 6000 Blackwell Server Edition 96GBFP8 (W8A8) | 1× · 126 tok/s | 2× · 33.6 tok/s |
| GeForce RTX 5090 32GBGGUF Q4_K_M | 2× · 138 tok/s | 2× · 45.7 tok/s |
| GeForce RTX 4090 24GBGGUF Q4_K_M | 4× · 85.6 tok/s | 4× · 37.0 tok/s |
| Apple M4 Max (40-core GPU, up to 128GB)GGUF Q4_K_M | 1× · 73.0 tok/s | 1× · 10.8 tok/s |
| DGX Spark (GB10, 128GB unified)FP8 (W8A8) | 1× · 25.2 tok/s | 1× · 3.3 tok/s |