neural.download

Home / Models / Nemotron 3.5 Lightning 30B-A3B

Nemotron 3.5 Lightning 30B-A3B

NVIDIA's Nemotron 3.5 Lightning, a 30B hybrid Mamba mixture-of-experts with about 3B active per word. A fast general assistant that runs on one Arc Pro B70 with stock software.

Nemotron 3.5 Lightning 30B-A3B · measured model
Open deployment packetAll packets and recipes
B70 fit
High
Quality evidence
Scoped deployment evidence
Interest
107.16K
Measured results
3

Packets and recipes

The deployment variants of this family, at every maturity.

nemotron-3.5-lightning-30b-a3bB70-verified

Nemotron 3.5 Lightning 30B-A3B · UD-Q4_K_M · one B70

Evidence packet · candidate package

decodeprefillcontext through 32Kqualityrecipe

Measured results

Every number links to its proof.

family data

One card vs. a two-card layer split

Same UD-Q4_K_M serving protocol; the two-card layer split is measured, not estimated

serving median · B70 cards 172.04–72.17tok/sDecodelab measurement; quality scope in evidence · evidence
serving median · B70 cards 269.45–69.49tok/sDecodelab measurement; quality scope in evidence · evidence
serving median
evidence 1

Speed as the conversation grows

One B70 · UD-Q4_K_M · f16 KV · raw llama-bench; native context beyond 32K remains unmeasured

020.440.961.381.802K4K8K16K24K32Kactive context tokens · tok/sUD-Q4_K_M · Decode (decode_tok_s) · active context tokens=0.0 · value=73.008586 tok/sUD-Q4_K_M · Decode (decode_tok_s) · active context tokens=2048.0 · value=72.420657 tok/sUD-Q4_K_M · Decode (decode_tok_s) · active context tokens=4096.0 · value=71.846756 tok/sUD-Q4_K_M · Decode (decode_tok_s) · active context tokens=8192.0 · value=70.6952 tok/sUD-Q4_K_M · Decode (decode_tok_s) · active context tokens=16384.0 · value=68.557408 tok/sUD-Q4_K_M · Decode (decode_tok_s) · active context tokens=24576.0 · value=66.552475 tok/sUD-Q4_K_M · Decode (decode_tok_s) · active context tokens=32768.0 · value=64.622975 tok/s
UD-Q4_K_M
UD-Q4_K_M 0: 73.01 · 2K: 72.42 · 4K: 71.85 · 8K: 70.7 · 16K: 68.56 · 24K: 66.55 · 32K: 64.62
evidence 1
Fine print

nemotron_h_moe, 53 layers, 2688 hidden. Only one model artifact and quantization are measured; no cross-quantization, cross-runtime, or later-weight transfer is claimed. Measurements, artifact hashes, outputs, quality decisions, and speed stay pinned to their exact recorded identity.

1 combination the lab has stopped pursuing: The artifact declares a 1,048,576-token native window, but this package has measured context-depth evidence only through 32,768 tokens. evidence Exact selectors are in the family data.

Many people at once Not measured

Multi-user (aggregate) throughput has not been measured for this family. The lab’s one measured sweep so far is in the multi-user report; the projection block below includes a projected users curve where a projection exists.

How much faster could this get? No projection

No projection is shown: the family has no curated headline measurement yet. The measured numbers above stand on their own.

Keep going

RecipesFilter runnable packetsChoose a modelQuality and deployment trade-offsHardwareCards, memory, and topologyFamily dataExact normalized coverage source