FinCrime-Atlas-TM-30B-A3B-Preview: an open-weight model for transaction monitoring investigations
Today we release FinCrime-Atlas-TM-30B-A3B-Preview (Atlas-TM), a model trained to investigate transaction monitoring alerts, make a filing decision and write a case memo. The model is a fine-tuned version of NVIDIA Nemotron 3.5 Lightning, a mixture-of-experts model with 30B parameters, of which 3B are active per token. We trained the model on full investigations by a teacher model in our case environment, with synthetic case data built on real-world financial crime scenarios. On FinCrime Atlas TM-Bench 1.0, Atlas-TM makes 7.5 percentage points more correct filing decisions than Haiku 5.5 (82.5% against 75.0%), with a TM-Bench 1.0 score 10.9 points lower (44.1 against 55.0). This release is a preview from checkpoint 3, and training continues.
Results
We report two measures on FinCrime Atlas TM-Bench 1.0, a set of 80 transaction monitoring cases at a US bank. Filing accuracy measures how often the model reaches the correct filing decision across all 80 cases. The TM-Bench 1.0 score combines decision correctness and memo quality into a score out of 100. A wrong or missing decision receives zero. For correct decisions, a model judge grades the memo against a private answer key, assessing factual coverage, evidence support and the explanation of the decision. We average these scores across all 80 cases.
Filing accuracy, % of cases
TM-Bench 1.0 score, out of 100
Estimated price for the 80 cases, based on actual token usage
Estimated price: the tokens each model used on the 80 cases, at the standard list price per million tokens, with no prompt caching discount. For FinCrime-Atlas-TM-30B-A3B-Preview, we use the OpenRouter price of its base model, Nemotron 3.5 Lightning ($0.07 input, $0.20 output). Haiku 5.5: the Anthropic price ($0.10 input and $0.50 output for prompts up to 100,000 tokens; $0.50 and $2.50 above).
What the model does
Atlas-TM works a transaction monitoring case from start to finish, the way a BSA/AML investigator does:
- Receive the case. Each case starts with one or more transaction monitoring alerts on a customer. With the alerts, the model receives the bank's procedures and policies, which set the standard for the review.
- Investigate. The model uses tools to review the customer profile and the transaction activity. The model chooses which records to examine and in which order, and follows up on what each record shows, as an investigator does.
- Decide. The model weighs the facts against the bank's SAR policy and decides whether the activity requires a filing: File SAR or No SAR.
- Document. The model records the investigation in a case memo, as the bank's procedures require. The memo explains the decision and ties each finding to the case records. When the decision is File SAR, the model also drafts the SAR narrative for the filing.
- Receive the caseAlerts
Procedures and policies - InvestigateCustomer profile
Transaction activity - DecideFile SAR
No SAR - DocumentCase memo
SAR narrative (File SAR)
How we trained it
FinCrime-Atlas-TM-30B-A3B-Preview is NVIDIA Nemotron 3.5 Lightning 30B-A3B, an open-weight mixture-of-experts model with 3B active parameters, fine-tuned with LoRA on full investigations in our environment. We trained three checkpoints. Each started from the base model, and each used more investigations than the one before. The model, the tools and the judge stayed the same. The score went from 7.2 to 44.1, and filing accuracy went from 51.3% to 82.5%. We release checkpoint 3, the latest checkpoint.
Filing accuracy, % of cases
TM-Bench 1.0 score, out of 100
Base: Nemotron 3.5 Lightning before training. Checkpoint 3 is FinCrime-Atlas-TM-30B-A3B-Preview. We assessed each model on the same 80 TM-Bench 1.0 cases. These cases were never used for training.
| Base model | nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16: 30B parameters, 3B active per token, BF16 weights |
| Method | LoRA supervised fine-tuning: rank 32, alpha 32, learning rate 1e-4, 2 epochs per training run, merged into the base weights for release |
| Framework | SkyRL 0.4.0 with the Megatron backend, with tensor and expert parallelism across the GPUs |
| Training hardware | One cloud node with 8 NVIDIA H100 SXM GPUs (80 GB each). |
| Training data | Full investigations of training cases by a teacher model, DeepSeek v4.1 Flash: the tool calls, the tool results, the reasoning and the case memo. We kept an investigation only when its decision was right and its memo graded well. |
| Serving and scoring | vLLM on 8 NVIDIA H100 GPUs, with a 1-million-token context window and tensor and expert parallelism. Each case has a limit of 64,000 output tokens per model turn and 60 tool calls. |
What is next
We continue to train Atlas-TM, with the next checkpoints focused on three priorities:
- More accurate filing decisions. Atlas-TM currently makes the correct filing decision on 82.5% of cases. We aim to reduce both missed SARs and unnecessary filings.
- More accurate and complete memos. Memos should support every claim with case evidence and capture all relevant facts in each section. We are working to reduce unsupported claims and citations, the largest contributor to the score gap with Haiku 5.5, and fill gaps in factual coverage.
- More reliable completion. Some runs hit tool-call or output limits, or end without a decision. Reducing these failures is a priority alongside improving decision and memo quality.
FinCrime Atlas TM-Bench 1.0 will remain unchanged so results stay comparable across model checkpoints. FinCrime Atlas TM-Bench 2.0 is in development, covering the same alert types with harder cases. Future versions will increase the difficulty to expose weaknesses as models improve.
Get the model
Download FinCrime-Atlas-TM-30B-A3B-Preview on Hugging Face under the OpenMDW-1.1 licence, the licence of the base model. The model card lists the model settings and the SHA-256 hash of each released weight file.
For the reported TM-Bench 1.0 evaluation, we used a 1-million-token context window, temperature 1.0, top-p 0.95, a maximum of 64,000 output tokens per model turn and up to 60 tool calls per case.
For access to the evaluation cases and environment, talk to us about benchmark access.
Work with us
AI labs
Train and evaluate models with training data, RL environments and FinCrime Atlas TM-Bench 1.0.
Banks and compliance teams
Assess filing decisions, memo quality and the evidence behind each claim.