Yes — with 28.0 GB to spare
Nemotron 3.5 Lightning 30B-A3B at Q4_K_M fits your Radeon Pro W7900 entirely on the GPU at 8K context, at an estimated 112 tokens per second. There is room for its full 256K window.
Mamba-2 + MoE hybrid built for the execution layer of agents: only 6 attention blocks, so the KV cache is almost free. Weights, data and recipe all open.
What hardware do I need for Nemotron 3.5 Lightning 30B-A3B? →
The VRAM budget
Quantisation ladder
| Quant | Weights | Total @ 8K | Max context | Tok/s | Quality | Fit |
|---|---|---|---|---|---|---|
| Q8_0 | 31.3 GB | 31.9 GB | 256K | 64 | −0.1% ppl | Long context |
| Q6_K | 24.1 GB | 24.8 GB | 256K | 82 | −0.4% ppl | Long context |
| Q5_K_M | 20.9 GB | 21.5 GB | 256K | 95 | −0.8% ppl | Long context |
| Q4_K_M | 17.8 GB | 18.4 GB | 256K | 112 | −1.9% ppl | Recommended |
| Q3_K_M | 14.4 GB | 15.0 GB | 256K | 138 | −5.4% ppl | Long context |
| Q2_K | 12.3 GB | 13.0 GB | 256K | 161 | −15% ppl | Long context |
Quality is the published perplexity delta against f16 weights. Max context assumes an f16 KV cache; q8_0 roughly doubles it. Only 6 of its 52 blocks keep a per-token KV cache; the rest are linear-attention, Mamba or convolution blocks with a fixed-size state.
How to run it
$ ollama pull nemotron-3.5-lightning:30b-a3b $ OLLAMA_CONTEXT_LENGTH=8192 \ ollama run nemotron-3.5-lightning:30b-a3b
The default. One binary, a model registry, an OpenAI-compatible port. More on Ollama.