MiMo V2.6 Distill Qwen 9B VRAM requirements
MiMo V2.6 Distill Qwen 9B has 9.4B parameters. With an 8K-token context and one request it needs about 6.61 GB of GPU memory at Q4_K_M, 10.4 GB at FP8 and 20.1 GB at FP16/BF16. The published weights take 17.5 GB (BF16). The smallest setup here that holds it at Q4_K_M with an 8K context is the RTX 4060 8GB; on one 24 GB RTX 3090 / 4090 it runs with its full 256K-token context.
6.61 GB at Q4_K_M, 8K context, one request
- FP8
- 10.4 GB
- FP16 / BF16
- 20.1 GB
- Published weights
- 17.5 GB
- Smallest setup, Q4_K_M
- RTX 4060 8GB
Worked example: MiMo V2.6 Distill Qwen 9B with 32K tokens
Inputs: MiMo V2.6 Distill Qwen 9B · Q4_K_M weights · 32,768 tokens of context · one request · FP16 KV cache
- Weights (9.4B at Q4_K_M)
- 5.30 GB
- KV cache (32 KB per token)
- 1.00 GB
- Buffers and runtime (0.5 GB + 10%)
- 1.13 GB
- Total
- 7.43 GB
The smallest setup here that holds it is the RTX 4060 8GB, with about 581 MB to spare. Change the inputs in the calculator
What makes MiMo V2.6 Distill Qwen 9B's memory use different
24 of its 32 layers keep a fixed-size linear-attention state, so the cache grows by 32 KB per token where caching every layer would take 128 KB.
Per token of context it adds 32 KB of FP16 cache; Qwen3 8B (8.2B), the nearest-sized model here with plain full attention, adds 144 KB, so MiMo V2.6 Distill Qwen 9B needs 22% as much.
How much VRAM does MiMo V2.6 Distill Qwen 9B need?
MiMo V2.6 Distill Qwen 9B needs 6.61 GB at Q4_K_M, 11.0 GB at Q8_0 and 20.1 GB at FP16 with 8,192 tokens of context and one request. Each total below is the weights plus the FP16 KV cache and the runtime overhead (0.5 GB plus 10%), and each row opens the calculator with that setting. What the GGUF names mean.
| Precision | Weights | Total | Smallest setup |
|---|---|---|---|
| As published (BF16) | 17.5 GB | 20.1 GB | RTX 3090 (24 GB) |
| FP16 / BF16 | 17.5 GB | 20.1 GB | RTX 3090 (24 GB) |
| FP8 / INT8 | 8.76 GB | 10.4 GB | RTX 3060 12GB |
| INT4 (AWQ / GPTQ) | 4.66 GB | 5.90 GB | RTX 4060 8GB |
| GGUF Q8_0 | 9.31 GB | 11.0 GB | RTX 3060 12GB |
| GGUF Q6_K | 7.19 GB | 8.68 GB | RTX 3060 12GB |
| GGUF Q5_K_M | 6.21 GB | 7.61 GB | RTX 4060 8GB |
| GGUF Q4_K_M | 5.30 GB | 6.61 GB | RTX 4060 8GB |
| GGUF IQ4_XS | 4.77 GB | 6.02 GB | RTX 4060 8GB |
| GGUF Q3_K_M | 4.28 GB | 5.49 GB | RTX 4060 8GB |
| GGUF IQ3_XXS | 3.61 GB | 4.75 GB | RTX 4060 8GB |
| GGUF Q2_K | 3.67 GB | 4.81 GB | RTX 4060 8GB |
Fine-tuning MiMo V2.6 Distill Qwen 9B? MiMo V2.6 Distill Qwen 9B VRAM for LoRA, QLoRA and full training.
Run MiMo V2.6 Distill Qwen 9B with llama.cpp
llama-server -hf bartowski/MiMo-V2.6-Distill-Qwen-9B-GGUF:Q4_K_M -c 262144 -ngl 99MiMo-V2.6-Distill-Qwen-9B-Q4_K_M.gguf, 5.8 GB, from bartowski/
MiMo V2.6 Distill Qwen 9B at 8K, 32K, 128K and 256K (full) tokens of context
8 of its 32 layers use full attention and 24 are linear-attention layers with no growing cache. Each extra token of context adds 32 KB of FP16 cache per request. The last column is the smallest setup here, one card or a group, that holds it at Q4_K_M. How the KV cache works.
| Context | KV cache, FP16 | Total at Q4_K_M | Total at Q8_0 | Smallest setup, Q4_K_M |
|---|---|---|---|---|
| 8K tokens | 256 MB | 6.61 GB | 11.0 GB | RTX 4060 8GB |
| 32K tokens | 1.00 GB | 7.43 GB | 11.8 GB | RTX 4060 8GB |
| 128K tokens | 4.00 GB | 10.7 GB | 15.1 GB | RTX 3060 12GB |
| 256K tokens | 8.00 GB | 15.1 GB | 19.5 GB | RTX 4060 Ti 16GB |
How fast MiMo V2.6 Distill Qwen 9B writes
Tokens per second for one request with 8,192 tokens of context, estimated from memory bandwidth. A dash means it does not fit on one card. Try other settings in the speed calculator.
| Hardware | Bandwidth | Q4_K_M | FP8 |
|---|---|---|---|
| RTX 3060 12GB | 360 GB/s | 32–44 | 20–28 |
| RTX 4090 | 1,008 GB/s | 82–119 | 53–75 |
| RTX 5090 | 1,792 GB/s | 132–203 | 88–130 |
| M4 Max Mac (128 GB) | 546 GB/s | 47–66 | 30–41 |
| M3 Ultra Mac Studio (512 GB) | 819 GB/s | 68–98 | 44–62 |
| H100 SXM | 3,350 GB/s | 211–348 | 148–230 |
| H200 | 4,800 GB/s | 266–464 | 194–314 |
Longest context on one GPU
How many tokens of context MiMo V2.6 Distill Qwen 9B fits on a single card with one request, an FP16 KV cache and 0.5 GB left free. "Full" means the model's whole context window fits.
| GPU | Memory | Q4_K_M | Q8_0 | FP8 |
|---|---|---|---|---|
| RTX 3060 | 12 GB | 150K | 21K | 39K |
| RTX 4060 Ti 16GB | 16 GB | 256K (full) | 138K | 155K |
| RTX 3090 / 4090 | 24 GB | 256K (full) | 256K (full) | 256K (full) |
| RTX 5090 | 32 GB | 256K (full) | 256K (full) | 256K (full) |
| A100 40GB | 40 GB | 256K (full) | 256K (full) | 256K (full) |
| Mac, 64 GB unified memory | 48 GB | 256K (full) | 256K (full) | 256K (full) |
| L40S / RTX 6000 Ada | 48 GB | 256K (full) | 256K (full) | 256K (full) |
| A100 / H100 80GB | 80 GB | 256K (full) | 256K (full) | 256K (full) |
| Mac, 128 GB unified memory | 96 GB | 256K (full) | 256K (full) | 256K (full) |
| H200 | 141 GB | 256K (full) | 256K (full) | 256K (full) |
| B200 | 180 GB | 256K (full) | 256K (full) | 256K (full) |
Model details
- Parameters
- 9.4B (9,409,813,744)
- Layers
- 8 of its 32 layers use full attention and 24 are linear-attention layers with no growing cache
- Attention cache
- 4 KV heads × 256
- Context length
- 262,144 tokens
- Published weights
- 17.5 GB (BF16)
- On Hugging Face
- XiaomiMiMo/
MiMo-V2.6-Distill-Qwen-9B
Why this estimate looks this way
At Q4_K_M and an 8K-token context, MiMo V2.6 Distill Qwen 9B uses 5.30 GB for weights, 256 MB for its FP16 KV cache and 1.06 GB for estimated runtime overhead, totaling 6.61 GB. The overhead is a 0.5 GB base plus 10% of weights and cache.
Its attention layout matters for long context: 8 of its 32 layers use full attention and 24 are linear-attention layers with no growing cache. The FP16 cache grows by about 32 KB per additional token per request.
The architecture and published checkpoint size come from the model's config.json and weight files. These are estimates rather than measured peak memory; inference engines can reserve extra buffers or preallocate the full configured cache.
MiMo V2.6 Distill Qwen 9B next to similar-sized models
The three models here closest to it in parameter count, at Q4_K_M with 8,192 tokens of context.
| Model | Parameters | Total | KV per token | Smallest setup |
|---|---|---|---|---|
| MiMo V2.6 Distill Qwen 9B | 9.4B | 6.61 GB | 32 KB | RTX 4060 8GB |
| Ornith 1.5 9B | 9.7B | 6.76 GB | 32 KB | RTX 4060 8GB |
| Qwen3.5 9B | 9.7B | 6.76 GB | 32 KB | RTX 4060 8GB |
| ZDTaichu 5.0 9B | 9.8B | 6.85 GB | 32 KB | RTX 4060 8GB |
Badge for your model card
Paste it into a Hugging Face or GitHub README; it shows the Q4_K_M total above and links to this page. Q8_0 and FP16 badges.
Numbers read from the model files on Hugging Face on . Compare it with other models in the reproducible model and GPU report.