The evidence on NVIDIA's Nemotron 3 family does not mainly show a model that leads its field. It shows how much capability survives when a model is made cheaper to run. The family covers Ultra, Super, Lightning and the small Nano and Bolt models. It is measured against Gemma, Qwen and DeepSeek peers. The efficiency measures include 4-bit weights and activations, speculative decoding through multi-token prediction (MTP), and 4B-class models run at reduced precision.
Most low-precision results show small losses. A framework of this kind has to balance its competing claims, and here the losses are uneven. A few of them, in agentic and multilingual work, are large enough to matter commercially.
We should be plain about the foundation. Nearly all the figures come from a single source, and the reporting window is narrow, 20–23 October 2026. Some numbers appear to come from the same technical-report tables, so they are not independent confirmations. Where a figure has more than one source, the text says so. Where it does not, the figure should be read as asserted, not established.
Quantization: strong at the median, uneven at the tail
The most robust group of results concerns low-precision recovery. Lightning in W4A16 form recovers a median of 99.24% of its BF16 results 1. Verbosity falls by 3.63% 1. AA-Omniscience slips by 1.12 1 and Tau3 Banking by 0.41 1. SWE-bench Verified is unchanged at 52.20 1.
The weak points are specific. Terminal-Bench v2.1 drops 2.39 pp, which is 90.22% recovery 1. AA-LCR falls from 52.38 to 50.00 1.
Super in NVFP4 shows the same mix. MMLU-ProX is down 3.15 pp, reported at 95.90% 1, and SWE-bench Multilingual is down 0.72 1. Against that, Terminal-Bench 2.1 is up 2.11 and CritPt is up 0.85 1. Losses do not fall evenly across tasks, and some tasks gain.
Two further data points show what is promising and what is unresolved.
- Promising: a pre-release candidate reached a 97.68% median under PTQ at 4.8 BPE 1. The released Super checkpoint's whole-model BPE is 4.82 1.
- Unresolved: failure behaviour degraded under quantization before a later fix. The Tau3-banking runaway rate was 28.2% for pre-MOPD FP4, against 19.7% for the BF16 reference 1. The "pre-MOPD" label implies a later stage, but the claims do not say how much that stage recovered.
There are two limits on how far these results extend. The authors state that W16A4 was not tested 1. The activation-only 4-bit case therefore remains unexamined, and the median recoveries describe the configurations tested, not low precision in general. Baselines are also not perfectly stable across the paper's tables. Super's BF16 MMLU-ProX baseline is reported as 76.91 in one table and 76.92 in another 1. The gap is trivial, but it is the kind of inconsistency that makes careful readers check the rest.
Speculative decoding: a real gain, but not a lead
MTP is the second efficiency lever, and the evidence for it is internal to NVIDIA's own variants. The SFT-stage MTP model reached an acceptance length of 4.528 at temperature 0 with draft length 7 1. Boosted Super reached 4.612 at temperature 1 on SPEED-Bench 1. The boosted variant improved acceptance at every draft position 1.
The competitive reading is less clear. Gemma-4 31B and Qwen3.5-122B-A10B matched or beat boosted Super in some reported columns 1. The improvement is real, but the claims do not show it giving NVIDIA a decisive edge over rivals.
Small models: Bolt's niche is instruction, role and grounding
The Bolt-E4B claims are the most heavily corroborated set, and they show the clearest pattern. In the BF16 reasoning table, Bolt leads Gemma4-E4B-it, Qwen3.5-4B and Nemotron-3 Nano-4B on instruction-following, role and grounding tasks.
- IFBench: Bolt scores 63.79, against 42.84 for Gemma and 58.91 for Qwen 2. Nano-4B scores 60.10 2.
- RoleMRC: Bolt scores 80.63, against 74.62 for Gemma and 65.81 for Qwen 2.
- RoleBench Custom: Bolt scores 67.05, against 53.66 for Qwen and 49.95 for Gemma 2. Nano-4B scores only 12.02, a figure confirmed by three sources 2.
- GaRAGe: Bolt scores 42.61, against 34.73 for Gemma and 24.66 for Qwen 2.
The picture reverses on tool use, long context and some other categories.
- Tau2-Bench: Qwen3.5-4B leads 75.06 to 55.10 2.
- BFCL-v3: Qwen scores 70.95 against Bolt's 68.51 2.
- LongBench v2: Qwen scores 59.93, against Bolt's 50.50 and Gemma's 49.34 2.
- RULER 64K: Qwen scores 95.15, against Bolt's 86.99 and Gemma's 86.55 2.
- EQ-Bench v3: Gemma leads 41.17 to 39.40 2.
- Vision: Bolt reportedly trails Gemma on MMMU-Pro Vision (42.02 vs 47.63), MathVision (46.25 vs 48.29) and MedXpertQA Vision (26.85 vs 28.20) 2.
The 4-bit builds carry the strongest result for Bolt. Its imatrix Q4_K_M build beats Gemma's Unsloth Q4_K_M on MultiChallenge (55.60 vs 39.36), RoleBench (71.21 vs 60.51) and IHEval (82.40 vs 62.87) 2. The same tension holds at Q4. Bolt loses to Gemma on BFCL-v3 (67.31 vs 70.56) 2. It also trails a Qwen3.5-4B Q4 score on LongBench v2 (48.67 vs 58.28) 2.
One caution applies to this whole section. Some figures are identical across the reasoning and non-reasoning tables. Bolt's 86.99 and Gemma's 86.55 on RULER appear in both 2. Nano's 19.31 on EQ-Bench also appears in both 2. This may be coincidence or reuse, and it should be verified before the numbers are treated as independent.
The frontier: Ultra is credible, not dominant
Ultra's comparisons with Qwen3.5-122B-A10B are mostly favourable.
| Benchmark | Ultra | Qwen3.5-122B-A10B |
|---|---|---|
| SWE-bench Multilingual | 68.78 1 | 62.00 1 |
| SWE-bench Pro | 49.66 1 | 45.60 1 |
| Apex-Shortlist | 75.00 1 | 54.65 1 |
| LiveCodeBench v6 | 89.21 1 | 76.65 1 |
| AA-LCR | 75.19 1 | 75.13 1 |
AA-LCR is effectively a tie. Ultra's other reported results include Terminal-Bench 2.1 at 54.78 1, BrowseComp at 67.85 1, CharXiv-RQ at 83.47 1 and HLE without tools at 29.24 1.
The record also contains two tensions on SWE-bench Pro. DeepSeek's V4 Flash-0731 scores 56.91, above Ultra 1. Super scores 51.78, also above Ultra's 49.66 1. The claims do not explain why the larger model trails the smaller one on this benchmark, and the gap should not be glossed over.
Super's training record shows a related pattern. Under an identical 33B-token SFT, mid-training moved SWE-bench Multilingual from 65.9 to 65.6 and SWE-bench Pro from 52.1 to 51.7. It raised SWE-bench Verified with the Claude Code harness from 53.8 to 58.0 1. The gain was concentrated in one agentic harness measure. Decontamination checks are also partial. The judge's recall on a synthetic Terminal-Bench 2.1 test is 87.5% 1.
What this implies
The following points are interpretation. They rest on the figures above and are not forecasts.
Quantization looks deployable for the cases tested, with identifiable weak points. Median recoveries of roughly 97–99% 1 suggest 4-bit variants can ship without large aggregate losses. The exceptions cluster in two places: long-horizon agentic work and multilingual breadth. Terminal-Bench v2.1 under Lightning W4A16 1, MMLU-ProX under Super NVFP4 1 and the pre-MOPD runaway rate 1 all point there.
Bolt's niche is instruction, role-play and grounding, not tool-use depth or long context. This is a reading of the benchmark pattern, not a statement about real-world workloads. Qwen3.5-4B is stronger on Tau2-Bench, BFCL-v3 and LongBench v2 2.
At the frontier, leadership is not established. Ultra beats a mid-sized open peer on most listed items. It trails DeepSeek V4 Flash on SWE-bench Pro, and Super outscores it there as well 1. This is the least resolved finding in the material.
Taken together, the family's strength is that the efficiency measures cost little at the median. Its weakness is that the costs fall on the tasks where reliability matters most. MTP acceptance lengths of 4.5–4.6 at draft length 7 1 are a real gain, but rivals match or exceed them in some columns 1. Given how much of the record rests on single sources and repeated tables, the numbers should be verified before anyone treats them as settled.
