NVIDIA Nemotron 3.5 Super VL is a 120B-parameter hybrid Mamba-Transformer mixture-of-experts model with 12B active parameters. Before its release, we fine-tuned the early-access checkpoint for finance and healthcare, and ran reinforcement learning on math, all through the Fastino Training API with a 75M-parameter LoRA adapter. These results use an early-access checkpoint. 

Two things stood out. Fine-tuning made Super better on every benchmark we measured, and it made Super cheaper to run: the fine-tuned model writes far fewer tokens while getting more answers right. This post covers the results, how fine-tuning cuts Super's output tokens, six lessons about fine-tuning Super, the tips we would give anyone starting today, and the recipes we would reuse.

Results at a glance

• Every benchmark improved. Two adapters, one per domain, raised all 15 held-out finance and healthcare benchmarks we measured, including five benchmarks they never optimized for.

• Finance (best of the two learning rates per benchmark; both tables below): FinQA 35.2% → 77.9%, TAT-QA 68.5 → 87.8 F1, SEC-Num 79.2% → 93.9%, FinEntity 84.7 → 92.4, and ConvFinQA, which contributed no training data, 22.3% → 76.3%. On four benchmarks the single combined adapter beats the single-task specialists we built for Nemotron 3.5 Lightning.

• Healthcare (best of two learning rates): BC5CDR 63.8 → 84.8, BioRED (never optimized for) 46.1 → 69.0, MedMentions 40.4 → 50.4, PubMedQA 70.2% → 78.6%, MEDEC 72.0% → 79.2%, MedCalc 65.9% → 72.5%.

• Cheaper to run: on PubMedQA the fine-tuned model answers in 113 tokens instead of 1,597 (14x fewer) and is 8.4 points more accurate; RL on math cut output tokens by 25% while adding 6 points.

• Little data: 500 FinQA rows, 15 minutes, and about $2 of training took FinQA from 35% to 77%.

• Reinforcement learning: MATH-500 84.5% → 91.4% (85.8% → 90.5% on a second seed) in 40 steps, with answers 25% shorter.

• Small and cheap: every adapter trains 0.06% of the model's parameters. A full domain adapter took one run of 2 to 5.5 hours on one B200 and $21 to $43 of training tokens at list price.

Finance

Super's base already reads financial tables well (TAT-QA 68.5 F1, against 17.2 for Lightning's base) but writes a correct executable FinQA program only a third of the time. The adapter teaches the program interface and accuracy more than doubles. The skill is general rather than memorized: ConvFinQA, a conversational benchmark that uses the same program language but contributed no training rows, rose from 22.3% to 76.3%, and FiQA and EDGARCalcQA, also untrained, rose 23.1 and 6.4 points. One combined adapter beats the Lightning specialists that each spent a whole rank-32 adapter on a single task: FinQA 77.9 vs 71.3, TAT-QA 87.8 vs 86.8, SEC-Num 93.9 vs 90.7, FinEntity 92.4 vs 85.1.

Healthcare

Extraction moved the most. Exact-span metrics give no partial credit, so every point is a correctly bounded, correctly typed entity. BC5CDR was only a small, capped slice of the healthcare mixture, yet the combined adapter beats the dedicated BC5CDR specialist we built for Lightning (84.8 vs 83.6), and the gain carried to BioRED, a different biomedical corpus the adapter never saw. MedMentions, which spans a far wider set of UMLS concept types, also passed Lightning's dedicated specialist at the higher learning rate (50.4 vs 49.7).

How little data does Super need?

We trained FinQA-only adapters on random subsets of the FinQA training set (question → gold program, rank 32, lr 3e-4, 3 epochs):

Nemotron 3.5 Lightning specialist, trained on all 6,251 FinQA examples (FinQA reference only).

Five hundred rows, fifteen minutes and about two dollars of training took Super from 35% to 77% on FinQA, past the Lightning specialist trained on every available row, and the skill carried to ConvFinQA, which received no training data. The last column is the warning: a single-skill adapter pulls a neighboring task with a different answer format toward its own format, and more single-task data pulls harder. The combined finance mixture, which includes TAT-QA's own format, raised TAT-QA instead (68.5 → 87.8).

Reinforcement learning

We ran the stock Tinker cookbook math_rl recipe, unchanged, against the Fastino Training API.

Base Super already solves 94% to 96% of the training problems, so the reward signal is thin. RL still added about 6 percentage points on average on held-out problems and cut output length by a quarter, on both seeds.

Better and cheaper: fewer tokens, more right answers

Super 3.5 is a strong reasoner, and a thorough one: it writes about 2.5x as many tokens as Nemotron 3 Super on the same GSM8K prompts. For teams serving it at scale, output tokens are the bill. Fine-tuning turned out to cut that bill and raise accuracy at the same time.

The healthcare adapter learned from short, correct rationales, and it learned their length along with their content. RL did the same on math: rewarded only for correct answers that finish within a 4,096-token budget, Super got 6 points more accurate and 25% shorter in 40 steps (table above).

• Serving cost scales with output tokens. A 14x cut on a classification-style task, or 25% on open-ended math, carries straight through to throughput and cost per request, on any inference stack.

• Accuracy and length are not a trade-off here. In every case the shorter model was also the more accurate one. On math, much of the gain came from answers that now finish within budget.

• It takes one run. The healthcare adapter trained in about 2 hours on one B200, about $21 at the Fastino Training API's list price.

Six lessons for fine-tuning Nemotron 3.5 Super

1. Teach the interface; Super already has the knowledge. Super's base scores are well above Lightning's on most tasks we measured (TAT-QA 68.5 vs 17.2, MedCalc 65.9 vs 48.6, MEDEC 72.0 vs 53.3, BC5CDR 63.8 vs 50.7). Its gaps are mostly interfaces: an executable program, an exact span with the right type, a flag in the right field. Those are exactly what small, clean datasets teach, and they produced the largest gains (FinQA +43, ConvFinQA +54, BC5CDR +21).

2. A few hundred correctly labeled training examples go a long way. 500 FinQA rows moved FinQA by 42 points in 15 minutes. Spend effort on the format and correctness of each row before spending it on volume.

3. Mix skills to protect neighbors. A narrow adapter drags nearby tasks toward its own output format (TAT-QA fell 5.5 points with 500 FinQA-only rows and 17.7 with 1,500). A balanced mixture that includes each format raised every task instead. If the model must keep several behaviors, put each one in the mixture.

4. Recipes transferred to Nemotron Super. Data and hyperparameters found on Lightning worked on Super unchanged, on every benchmark we measured. The expensive part of fine-tuning is the search; on Nemotron it can be done on the 30B-A3B model and replayed on the 120B-A12B model in one run.

5. Fine-tuning teaches length, not just content. Super 3.5 reasons at length: 1,597 tokens per PubMedQA answer with thinking on. Trained on concise, correct rationales, it answers in 113 tokens and is 8.4 points more accurate. RL on math did the same (25% shorter, 6 points better). If inference cost matters, supervise the length you want.

6. Use 1x to 2x the full-adapter learning rate, and give the model room to think. The default Super adapter leaves the routed experts frozen and trains 75M parameters. At 2x the original rate, healthcare gains grew on seven of eight benchmarks (MedCalc +6.5 vs +2.4, EkaCare +3.8 vs +0.4); in finance both rates landed within 1.6 points on the trained tasks, and 2x transferred better (ConvFinQA +54.0 vs +49.3, EDGARCalcQA +6.4 vs +2.7). And budget tokens generously: in an earlier check with a 1,024-token limit, base Super scored 60.2% on MATH-500 because 39% of answers were cut off; with 4,096 tokens it scores 84.5% to 85.8%.

Tips

• Supervise the artifact. Programs, spans and labels as targets, with no explanatory prose, transferred cleanly and gave the largest gains.

• Select on held-out behavior, not training loss, and keep two scoreboards: the target benchmark and an untouched neighbor (BioRED for BC5CDR, ConvFinQA for FinQA). Transfer tells you whether the model learned the skill or the surface.

• Keep the template identical between training and serving, including the thinking flag. Super's chat template renders reasoning and non-reasoning turns differently.

• Watch tool-call format. The base model sometimes emits several JSON tool calls in one message; a small amount of format-matched SFT or RL fixes it quickly.

• Train and sample long when you need to. Super trains and samples at up to 64K tokens on the Fastino Training API.

Recipes we would reuse on Super

Cost

Appendix: protocol

• Early-access checkpoint, text inputs, Super 3.5's own chat template; thinking off except PubMedQA, where the Lightning protocol had it on.

• Temperature 0 for all SFT evaluations; RL evaluation at temperature 1 with 4 samples per problem. One seed for SFT, two for RL.

• Max tokens raised above the Lightning protocol on free-text tasks to fit Super's longer answers; truncation stayed under 2% on every SFT benchmark.

• BizFinBench: the six objectively scored tasks. FinEntity and FiQA: FinBERT tokenizer span alignment.

• Small sets (FinEntity n=197, BioRED 66 documents) carry 1 to 2 points of noise.