Nemotron 3.5 Lightning 30B-A3B
NVIDIA's compact hybrid MoE combines 6 attention blocks with 23 Mamba-2 blocks and 23 expert blocks. It activates about 3B of 30B language-model parameters, supports up to a one-million-token context, and has official local-AI recipes for RTX workstations and DGX Spark.
NVIDIAMixture of expertsagenticreasoning26 machines can run it
Total parameters
31.578B
Official spec sets memory need
Active parameters
3B
Official spec sets decode speed
Context
1M
Official spec tokens
4-bit weights
17 GB
Calculated before KV cache
This is a sparse mixture of experts. All 31.578B of weights must be available to the runtime, but only 3B are read per generated token. Ordinary runtimes keep the full quantized model in memory; a runtime-specific SSD-streaming artifact can retain a smaller working set and fetch expert data on demand, trading speed for capacity.
Model specification
Publisher
NVIDIA Official spec
Architecture
Sparse mixture of experts Official spec
Total parameters
31.578B Official spec
Active parameters
3B per token Official spec
Context length
1,048,576 tokens Official spec
Attention ⓘ
Hybrid Mamba-2 · 6 attention + 23 Mamba blocks Official spec
Licence
OpenMDW 1.1 Official spec source
Specification confidence ⓘ
High · verified 2026-09-21 Official spec source
Released
11 Aug 2026 Official spec
Official source
Quantizations and memory
| Quantization | Format | Bits/weight | Resident | Download | Quality kept |
|---|---|---|---|---|---|
MLX 4-bit | mlx | 4.5 | 17.2 GB | 17.2 GB | 98.0% |
Q4_K_Mdefault | gguf | 4.85 | 18.4 GB | 18.4 GB | 98.5% |
Q5_K_M | gguf | 5.69 | 21.5 GB | 21.5 GB | 99.3% |
Q8_0 | gguf | 8.5 | 31.9 GB | 31.9 GB | 99.9% |
BF16 | safetensors | 16 | 58.8 GB | 58.8 GB | 100.0% |
Generic weight sizes are computed from the parameter count and bits per weight plus a format-specific overhead. Runtime-specific artifacts use their published resident and download footprints; streamed models can therefore require much more disk than memory. Quality retention is an assumption, not a measured evaluation.
Coding & quality benchmarksCompare coding results →
No published benchmark results have been imported for this model yet. This is missing evidence, not a score of zero.
We show source metrics rather than deriving one opaque quality number. Different benchmarks measure genuinely different things, and collapsing them into a single score would hide exactly the disagreements worth seeing.
Recommendations
Cheapest that can run it
~9.7–14 t/s · $1,531
Lowest purchase price among configurations where the model fits at some quantization in our catalogue. Speed is not considered.
Cheapest above 20 t/s
~22.5–32.3 t/s · $1,711
20 tokens/second is roughly the point at which generation keeps pace with reading. Below it, interactive use feels like waiting.
Cheapest above 40 t/s
~213–307 t/s · $1,981
40 tokens/second is the threshold most people describe as comfortable for coding agents, where output arrives faster than you can review it.
Fastest with real measurements
Nothing in the database qualifies.
Highest throughput among configurations with an actual published measurement rather than our estimate.
Best throughput per purchase-price unit
~237–342 t/s · $2,071
Highest decode tokens/second per 1,000 units of the displayed purchase currency. Ignores running costs and resale — see the economics section for the full picture.
Hardware that runs Nemotron 3.5 Lightning 30B-A3B0 measured, 26 estimated
26 of 26 rows
| Hardware↕ | Quant | Memory↕ | Decode▼ | Prefill↕ | Context | Price↕ | Fit | Confidence↕ |
|---|---|---|---|---|---|---|---|---|
| NVIDIA DGX H200 (8x H200, 1,128 GB) NVIDIA · 1,128 GB · 38,400 GB/s | Q8_0recommended | 36.7 GB | ~533–768 t/s | ~274900–570940 t/s | 1024K | $359,430 | Comfortable | Estimated |
| RTX 5090 workstation (1x 32 GB) NVIDIA · 32 GB · 1,792 GB/s | Q5_K_Mrecommended | 23.3 GB | ~319–459 t/s | ~3170–6580 t/s | 512K | $3,332 | Comfortable | Estimated |
| Quad RTX 5090 workstation (4x 32 GB) NVIDIA · 128 GB · 7,168 GB/s | Q8_0recommended | 35.1 GB | ~315–454 t/s | ~10290–21380 t/s | 1024K | $12,611 | Comfortable | Estimated |
| Lenovo ThinkStation PX (4x RTX PRO 6000, 384 GB) Lenovo · 384 GB · 7,168 GB/s | Q8_0recommended | 35.1 GB | ~293–422 t/s | ~13200–27410 t/s | 1024K | $58,553 | Comfortable | Estimated |
| Dual RTX 5090 workstation (2x 32 GB) NVIDIA · 64 GB · 3,584 GB/s | Q8_0recommended | 34.3 GB | ~249–359 t/s | ~5150–10690 t/s | 1024K | $5,854 | Comfortable | Estimated |
| Used RTX 4090 workstation (24 GB) NVIDIA · 24 GB · 1,008 GB/s | Q4_K_Mrecommended | 20.0 GB | ~237–342 t/s | ~3530–7320 t/s | 192K | $2,071 | Borderline | Estimated |
| Dell Precision 7960 Rack (2x RTX PRO 6000, 192 GB) Dell · 192 GB · 3,584 GB/s | Q8_0recommended | 34.3 GB | ~236–339 t/s | ~6600–13700 t/s | 1024K | $36,663 | Comfortable | Estimated |
| RTX PRO 6000 Blackwell workstation (96 GB) NVIDIA · 96 GB · 1,792 GB/s | Q8_0recommended | 33.9 GB | ~227–326 t/s | ~3340–6940 t/s | 1024K | $10,359 | Comfortable | Estimated |
| RTX PRO 6000 Max-Q workstation (96 GB, 300 W) NVIDIA · 96 GB · 1,792 GB/s | Q8_0recommended | 33.9 GB | ~227–326 t/s | ~2740–5690 t/s | 1024K | $10,809 | Comfortable | Estimated |
| Dual used RTX 4090 workstation (48 GB) NVIDIA · 48 GB · 2,016 GB/s | Q5_K_Mrecommended | 23.7 GB | ~223–321 t/s | ~5730–11900 t/s | 1024K | $3,692 | Comfortable | Estimated |
| RTX 3090 workstation (Ryzen 9 7950X, 64 GB) NVIDIA · 24 GB · 936 GB/s | Q4_K_Mrecommended | 20.0 GB | ~213–307 t/s | ~1520–3150 t/s | 192K | $1,981 | Borderline | Estimated |
| NVIDIA DGX Station GB300 (748 GB) NVIDIA · 748 GB · 7,100 GB/s | Q8_0recommended | 33.9 GB | ~181–261 t/s | ~36730–76290 t/s | 1024K | $90,082 | Comfortable | Estimated |
| Mac Studio M5 Ultra 512 GB Apple · 512 GB · 1,200 GB/s | Q8_0recommended | 33.9 GB | ~180–258 t/s | ~2980–6190 t/s | 1024K | $11,350 | Comfortable | Estimated |
| Mac Studio M5 Ultra 256 GB Apple · 256 GB · 1,200 GB/s | Q8_0recommended | 33.9 GB | ~180–258 t/s | ~2980–6190 t/s | 1024K | $8,647 | Comfortable | Estimated |
| Mac Studio M5 Ultra 96 GB Apple · 96 GB · 1,200 GB/s | Q8_0recommended | 33.9 GB | ~180–258 t/s | ~2390–4950 t/s | 1024K | $5,945 | Comfortable | Estimated |
| Radeon AI PRO R9700 workstation (32 GB) AMD · 32 GB · 644 GB/s | Q5_K_Mrecommended | 23.3 GB | ~167–241 t/s | ~2530–5240 t/s | 512K | $2,521 | Comfortable | Estimated |
| Mac Studio M5 Max 36 GB Apple · 36 GB · 614 GB/s | Q5_K_Mrecommended | 23.3 GB | ~145–208 t/s | ~1190–2480 t/s | 1024K | $2,702 | Comfortable | Estimated |
| Mac Studio M3 Ultra 512 GB Apple · 512 GB · 819 GB/s | Q8_0recommended | 33.9 GB | ~110–158 t/s | ~719–1490 t/s | 1024K | $8,287 | Comfortable | Estimated |
| Mac Studio M3 Ultra 256 GB Apple · 256 GB · 819 GB/s | Q8_0recommended | 33.9 GB | ~110–158 t/s | ~719–1490 t/s | 1024K | $5,674 | Comfortable | Estimated |
| Mac Studio M5 Max 128 GB Apple · 128 GB · 614 GB/s | Q8_0recommended | 33.9 GB | ~103–148 t/s | ~1490–3100 t/s | 1024K | $4,323 | Comfortable | Estimated |
| Mac Studio M5 Max 64 GB Apple · 64 GB · 614 GB/s | Q8_0recommended | 33.9 GB | ~103–148 t/s | ~1490–3100 t/s | 1024K | $3,512 | Comfortable | Estimated |
| Mac mini M4 Pro 64 GB Apple · 64 GB · 273 GB/s | Q8_0recommended | 33.9 GB | ~39.8–57.2 t/s | ~125–259 t/s | 1024K | $2,251 | Comfortable | Estimated |
| NVIDIA DGX Spark 128 GB NVIDIA · 128 GB · 273 GB/s | Q8_0recommended | 33.9 GB | ~37.3–53.6 t/s | ~1280–2670 t/s | 1024K | $3,873 | Comfortable | Estimated |
| Framework Desktop (Ryzen AI Max+ 395, 128 GB) AMD · 128 GB · 256 GB/s | Q8_0recommended | 33.9 GB | ~25.7–37 t/s | ~355–737 t/s | 1024K | $2,161 | Comfortable | Estimated |
| GMKtec EVO-X2 (Ryzen AI Max+ 395, 128 GB) AMD · 128 GB · 256 GB/s | Q8_0recommended | 33.9 GB | ~22.5–32.3 t/s | ~355–737 t/s | 1024K | $1,711 | Comfortable | Estimated |
| CPU-only workstation (Ryzen 9950X, 192 GB DDR5) Generic · 192 GB · 90 GB/s | Q8_0recommended | 33.9 GB | ~9.7–14 t/s | ~42.7–88.8 t/s | 1024K | $1,531 | Comfortable | Estimated |
Submit a benchmarkContributions are reviewed before publication.
Similar models
Nemotron 3 Super 120B-A12B
NVIDIA · 123.611B (12B active) · NVIDIA Nemotron Open Model License
Qwen3.8 Flash Next
Alibaba Qwen · 180B (6B active) · Qwen Community 1.0
Qwen3-Coder 30B-A3B
Alibaba Qwen · 30.532B (3.3B active) · Apache 2.0
Qwen3 30B-A3B
Alibaba Qwen · 30.532B (3.3B active) · Apache 2.0
Step 3.7 Flash 198B-A11B
StepFun · 201.365B (11B active) · Apache 2.0
MiMo-V2.6 Flash 309B-A15B
XiaomiMiMo · 309B (15B active) · MIT