SAVRN Model Hub · Comparisons
NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4 vs Qwen3.6-35B-A3B-NVFP4
NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4 has 17.8B parameters and Qwen3.6-35B-A3B-NVFP4 has 18.7B parameters; NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4 is released under other and Qwen3.6-35B-A3B-NVFP4 under Apache License 2.0; at 16-bit, NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4 needs about 42.8 GB (1x MI300X from $1.85 an hour) and Qwen3.6-35B-A3B-NVFP4 about 44.8 GB (1x MI300X from $1.85 an hour).
| Field | NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4 nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4 | Qwen3.6-35B-A3B-NVFP4 nvidia/Qwen3.6-35B-A3B-NVFP4 |
|---|---|---|
| Publisher | NVIDIA | NVIDIA |
| Task | Text generation | Text generation |
| Modality | Text | Text |
| Parameters, as reported | 17.8B parameters | 18.7B parameters |
| Architecture | NemotronHForCausalLM | Qwen3_5MoeForConditionalGeneration |
| Library | transformers | Model Optimizer |
| Context length | 1,048,576 tokens | 262,144 tokens |
| Repository size | 21.6 GB | 23.5 GB |
| Artifact formats | safetensors, pytorch | safetensors |
| License | other | apache-2.0 |
| Access | Open weights, no gate | Open weights, no gate |
| Memory at 16-bit (weights and margin) | 42.8 GB | 44.8 GB |
| Cheapest GPUs at 16-bit, per hour | 1x MI300X, $1.85 | 1x MI300X, $1.85 |
| Memory at 4-bit (weights and margin) | 10.7 GB | 11.2 GB |
| Cheapest GPUs at 4-bit, per hour | 1x MI300X, $1.85 | 1x MI300X, $1.85 |
| Revision viewed | bee7596271d1 | 1355db6a0524 |
| Downloads reported by the hub | 1.2M | 4.6M |
| Last observed | 2026-09-19 | 2026-10-16 |
An evaluation row appears only where at least two of these models report the same benchmark with the same stated configuration, metric, unit and setup. Different evaluators stay named in each cell. Values are shown as reported: no unit conversion, no ranking.
Other Reported Results
These results are listed for each model on its own, because the conditions needed to compare them are not stated or do not match. Two results that leave a condition blank are not assumed to share it.
NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4
| Benchmark | Conditions | Result | Reported by | Revision | Date |
|---|---|---|---|---|---|
| Idavidrein/gpqa | Task diamondMetric diamondComparison conditions not established | 75.57 | NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4 model card Reported by a third party |
Evaluated revision not stated | 2026-08-13 |
| SWE-bench/SWE-bench_Multilingual | Task swe_bench_multilingual_%_resolvedMetric swe_bench_multilingual_%_resolvedComparison conditions not established | 36.47 | NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4 model card Reported by a third party |
Evaluated revision not stated | 2026-08-13 |
| SWE-bench/SWE-bench_Verified | Task swe_bench_%_resolvedMetric swe_bench_%_resolvedComparison conditions not established | 52.8 | NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4 model card Reported by a third party |
Evaluated revision not stated | 2026-08-13 |
| TIGER-Lab/MMLU-Pro | Task mmlu_proMetric mmlu_proComparison conditions not established | 81.62 | NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4 model card Reported by a third party |
Evaluated revision not stated | 2026-08-13 |
| cais/hle | Task hleMetric hleSetup Text-only, no tools — HLE's default includes image questions, so this excludes the multimodal subset.Comparison conditions not established | 10.47 | NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4 model card Reported by a third party |
Evaluated revision not stated | 2026-08-13 |
SAVRN's Notes on NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4
Despite the 30B-A3B name, the Index prices this at 17.8B parameters: 42.8 GB of working memory at 16-bit, 21.4 GB at 8-bit, and 10.7 GB at 4-bit, so one MI300X at $1.85 an hour covers every precision listed. The headline is the 1,048,576-token context — a full million tokens — which is where the memory actually goes once you push past short prompts.
The license is 'other' on our Index; read it before commercial use rather than assuming a permissive grant. It is a Nemotron-H mixture-of-experts with 6 active experts, quantized with NVIDIA's own Model Optimizer, and the card lists a September 2025 pretraining cutoff and a May 2026 post-training cutoff. Neighbors include the Nemotron-3-Nano-30B and Qwen3.6-35B-A3B quants. It draws 1.2M downloads a month.
SAVRN's Notes on Qwen3.6-35B-A3B-NVFP4
Read the artifact, not the table. NVIDIA's contribution to this release is the quantization, not the architecture: Alibaba's Qwen3.6-35B-A3B — 18.7B stored parameters as 256 experts with 8 active per token — run through Model Optimizer into three safetensors files totalling 23.4 GB. The generic memory table models a dense 18.7B, 44.8 GB at 16-bit and 11.2 GB at 4-bit; your actual download lands between those figures on one 192 GB MI300X at $1.85 an hour.
Apache 2.0 governs, and NVIDIA's card is explicit that it neither owns nor developed the model. The recipe is ANCFG: static per-block 4-bit weights at group size 16, dynamic 4-bit activations, FP8 attention, local-Hessian calibration. First published 2026-05-27, updated 2026-08-29. No Index host prices it per token, so the card hour is your cost basis — and the 262,144-token context is where the spare memory goes.
Questions
Which is larger, NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4 or Qwen3.6-35B-A3B-NVFP4?
Qwen3.6-35B-A3B-NVFP4 (18.7B parameters) is larger than NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4 (17.8B parameters), by the parameter counts their publishers report.
Which is cheaper to run, NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4 or Qwen3.6-35B-A3B-NVFP4?
At 4-bit, NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4 fits on 1x MI300X from $1.85 an hour and Qwen3.6-35B-A3B-NVFP4 on 1x MI300X from $1.85 an hour, at the lowest on-demand prices the SAVRN Index lists.
Can I use Qwen3.6-35B-A3B-NVFP4 commercially?
Yes. Qwen3.6-35B-A3B-NVFP4 is released under Apache License 2.0. The Apache License 2.0 is a permissive open-source license. It permits commercial use, modification and redistribution. It requires keeping the license and copyright notices and any NOTICE file, stating significant changes, and it includes an express patent grant from contributors.