Nemotron-3.5-Super-Law-VL-123B-A14B-NVFP4
Super-Law is built on NVIDIA Nemotron 3.5 Super and post-trained with 9 steps of GRPO on agentic legal work: multi-turn tasks where the model reads a matter's documents, uses tools and writes the deliverable. It scores 5–9% higher than the base model and improves in 21 of 25 practice areas on legal tasks it never saw. It is a standard NVFP4 checkpoint: stock vLLM serves it with no plugins or custom code.
Results
64 held-out tasks from the Harvey LAB legal-agent benchmark, across 25 practice areas, none of them seen in training. Each deliverable is graded against the task's rubric by an LLM judge; the score is the share of rubric criteria met.
| Base | Super-Law | Change | |
|---|---|---|---|
| All 64 held-out tasks | 0.841 | 0.886 | +5.4% |
| Held-out tasks, same judge model for both (51) | 0.834 | 0.888 | +6.5% |
| Held-out tasks in the practice areas it trained in, same judge (9) | 0.832 | 0.891 | +7.1% |
| Its training tasks, during training (best step) | 0.740 | 0.806 | +8.9% |
Base scores come from an earlier run of the base model on the same 64 tasks; 51 of its deliverables were re-graded with the judge model used for Super-Law.
Gains by practice area
| Practice area | Base | Super-Law | Change |
|---|---|---|---|
| Contracts | 0.748 | 0.947 | +26.5% |
| Emerging companies & venture capital | 0.776 | 0.924 | +19.1% |
| Capital markets | 0.800 | 0.881 | +10.1% |
| International trade & sanctions | 0.739 | 0.811 | +9.6% |
| Immigration | 0.819 | 0.885 | +8.2% |
| Banking & finance | 0.827 | 0.891 | +7.7% |
| Funds & asset management | 0.828 | 0.881 | +6.4% |
| Healthcare & life sciences | 0.878 | 0.925 | +5.4% |
| Litigation & dispute resolution | 0.780 | 0.818 | +4.8% |
| Antitrust & competition | 0.915 | 0.952 | +4.1% |
| Tax | 0.832 | 0.862 | +3.6% |
| Data privacy & cybersecurity | 0.914 | 0.942 | +3.1% |
| Trusts, estates & private client | 0.861 | 0.883 | +2.5% |
| White-collar defense & investigations | 0.814 | 0.834 | +2.4% |
| Corporate governance | 0.918 | 0.940 | +2.3% |
| Structured finance & securitization | 0.832 | 0.850 | +2.1% |
| Arbitration & international dispute resolution | 0.881 | 0.899 | +2.0% |
| Energy & natural resources | 0.928 | 0.946 | +2.0% |
| Insurance | 0.868 | 0.879 | +1.2% |
| Intellectual property | 0.834 | 0.844 | +1.2% |
| Real estate | 0.893 | 0.902 | +1.0% |
Usage
vllm serve LLMWildling/Nemotron-3.5-Super-Law-VL-123B-A14B-NVFP4 \
--tensor-parallel-size 2 \
--trust-remote-code \
--enable-auto-tool-choice \
--tool-call-parser qwen3_coder \
--reasoning-parser nemotron_v3
Tested with vLLM 0.32, no plugins. See the base model card for the reasoning parser and other serving options.
License
OpenMDW-1.1, inherited from NVIDIA Nemotron 3.5 Super VL.