Llama-3.3-Nemotron-Cascade-2-30B-A3B-Curator-0.1.0-Enrichment-Thinking-OFF
Curator Research Private Limited · Version 0.1.0 · Public experimental release
Highlights
Built with Llama: trained on Llama 3.3 70B outputs; architecture nvidia/Nemotron-Cascade-2-30B-A3B.
Structured extraction of entities, facts and citations from documents.
- Fastest single-stream decoding in Curator 0.1.0: 154 tokens/s (5.8 ms per token) on one A100, and 1,166 tokens/s at 32 concurrent requests.
- On 48 held-out documents: valid JSON on 40/48 (8 answered in prose); over all 48, schema 0.748 and grounding 0.636 (0.931 and 0.784 on the 40 parsed).
Assessment: Fast full-weight enrichment model with native vLLM support. 40/48 outputs parse as JSON, and the 8B enrichment adapters lead on grounding.
Architecture: nvidia/Nemotron-Cascade-2-30B-A3B · Artifact: full · Mode: thinking-off.
Family comparison
Evaluation pools and paired comparisons are described below. This release is shown in bold.
Enrichment evaluation — 1 October 2026
48 held-out documents per measured release.
| HF model | Valid JSON | Schema | Grounding | Semantic F1 |
|---|---|---|---|---|
| Llama-3.3-Qwen3-8B-Curator-0.1.0-Enrichment-Thinking-OFF | 48/48 | .945 | .858 | .499 |
| Llama-3.3-Qwen3-8B-Curator-0.1.0-Enrichment-Thinking-ON | 48/48 | .970 | .877 | .559 |
| Llama-3.3-Qwen3.6-35B-A3B-Curator-0.1.0-Enrichment-Thinking-OFF | 48/48 | .985 | .753 | .398 |
| Llama-3.3-Qwen3.6-35B-A3B-Curator-0.1.0-Enrichment-Thinking-ON | 36/48 | .741 | .659 | .214* |
| Llama-3.3-Nemotron-Cascade-2-30B-A3B-Curator-0.1.0-Enrichment-Thinking-OFF | 40/48 | .748 | .636 | .399 |
| Llama-3.3-Nemotron-Cascade-2-30B-A3B-Curator-0.1.0-Enrichment-Thinking-ON | Pending | — | — | — |
*35B ON semantic F1 covers 47 scored documents; one was unscored. Nemotron ON is not yet evaluated: no held-out case could be proven unseen by this checkpoint.
Paired comparison — 8 documents per release
Values show student / Llama 3.3 70B teacher.
| HF model | Schema | Semantic F1 | Judge |
|---|---|---|---|
| Llama-3.3-Qwen3-8B-Curator-0.1.0-Enrichment-Thinking-ON | .989 / .933 | .508 / .571 | .750 / .781 |
| Llama-3.3-Qwen3-8B-Curator-0.1.0-Enrichment-Thinking-OFF | .915 / .872 | .527 / .572 | .727 / .852 |
| Llama-3.3-Nemotron-Cascade-2-30B-A3B-Curator-0.1.0-Enrichment-Thinking-OFF | .708 / .872 | .396 / .572 | .594 / .852 |
| Llama-3.3-Qwen3.6-35B-A3B-Curator-0.1.0-Enrichment-Thinking-OFF | .979 / .872 | .267 / .572 | .461 / .852 |
| Llama-3.3-Qwen3.6-35B-A3B-Curator-0.1.0-Enrichment-Thinking-ON | .750 / .933 | .150 / .571 | .406 / .781 |
The eight-document sample limits generalisation. Schema measures format conformance. References and judge use the teacher family, which can favour the teacher.
Checkpoint-selection comparison (training record)
Base and selected adapter used the same 120-document selection slice; run dates were not preserved.
| HF model | Quality: base → adapter | Semantic F1: base → adapter | Grounding: base → adapter |
|---|---|---|---|
| Llama-3.3-Qwen3-8B-Curator-0.1.0-Enrichment-Thinking-OFF | .610 → .716 (+17%) | .344 → .583 (+69%) | .994 → .889 |
| Llama-3.3-Qwen3-8B-Curator-0.1.0-Enrichment-Thinking-ON | .610 → .736 (+21%) | .344 → .629 (+83%) | .994 → .839 |
| Llama-3.3-Qwen3.6-35B-A3B-Curator-0.1.0-Enrichment-Thinking-ON | .670 → .723 (+8%) | .468 → .589 (+26%) | .956 → .874 |
| Llama-3.3-Qwen3.6-35B-A3B-Curator-0.1.0-Enrichment-Thinking-OFF | .670 → .678 (+1%) | .468 → .560 (+20%) | .956 → .787 |
Percentages are relative changes calculated from rounded scores. The 35B OFF checkpoint is recorded as a baseline selection. A 3 October check shows the adapter does change the base: on a fixed prompt the base's most likely first token ends the reply, while the adapter's opens the JSON answer.
GPU serving — 1 October 2026
One NVIDIA A100 80 GB, vLLM 0.29. Each row describes a shared serving package.
| Package / representative HF model | Base weights | Single stream | Concurrent throughput | GPU J/token | KV capacity (tokens) |
|---|---|---|---|---|---|
| Llama-3.3-Qwen3-8B-Curator-0.1.0-Enrichment-Thinking-ON + 6 sibling adapters (AWQ) | 6.1 GB | 111 tok/s | 684 tok/s at 32 | .56 | 471,200 |
| Llama-3.3-Qwen3.6-35B-A3B-Curator-0.1.0-Enrichment-Thinking-ON + 3 sibling adapters (W4A16) | 21.8 GB | 94 tok/s | 1,088 tok/s at 32 | .30 | 1,929,911 |
| Llama-3.3-Nemotron-Cascade-2-30B-A3B-Curator-0.1.0-Enrichment-Thinking-OFF (BF16) | 63.2 GB | 154 tok/s | 1,166 tok/s at 32 | .33 | 1,358,506 |
| Llama 3.3 70B AWQ | 39.8 GB | — | 107 tok/s at 16, warm | 3.70 | 108,976 |
Student sweeps used 2,048 input / 256 output tokens and cold cache. The teacher figure comes from a 180-second warm window with different output lengths. Ratios between these rows combine model and workload effects. Energy covers sampled GPU board power. The 8B base occupies about one-sixth of the teacher's weight storage and serves seven adapters. Nemotron ON throughput remains unmeasured.
Selection measurements (training record)
Recorded result: validation loss 0.0698 (lower is better; not a task-quality score).
Comparison pool: Curator enrichment validation split.
From the training record; run dates were not preserved.
Loading and use
Load the full model using a runtime compatible with nvidia/Nemotron-Cascade-2-30B-A3B. The measured weight files match the published weight files; the immutable public loading pin is shown below.
Native NemotronH serving uses vLLM 0.29, with remote-code execution disabled:
vllm serve curator-research/Llama-3.3-Nemotron-Cascade-2-30B-A3B-Curator-0.1.0-Enrichment-Thinking-OFF --revision 29fd1d739c5970a433920c4e68ca10a064cc8769 --dtype bfloat16 \
--served-model-name Llama-3.3-Nemotron-Cascade-2-30B-A3B-Curator-0.1.0-Enrichment-Thinking-OFF
OFF was measured in this runtime; ON has no held-out serving result yet.
Set the chat template's enable_thinking option to false for this variant. The bundled template thinks by default, and the served model name does not change it. The 1 October results used Curator's training-time prompt format; this chat-template setting was not separately evaluated.
Curator publishes these weights as an experimental release of its applied model research, with its own measurements on held-out data. Training data, prompts and the output schema remain private, so the published scores cannot be reproduced outside Curator; evaluate this exact revision on your own data before production use.
License and lineage
The NVIDIA Open Model License applies to the merged weights and includes its Trustworthy AI conditions (LICENSE, NOTICE.txt). Because its training targets came from Llama 3.3 70B, the Llama 3.3 Community License attribution and naming terms and the Llama Acceptable Use Policy also apply (TEACHER_LICENSE.txt). Both apply together; this is not a dual licence.
Curator Research
Read the model overview and evaluation boundaries, explore the Curator Research organisation, and follow Curator Research on LinkedIn.