Introducing Vulcan Mini: Our first model optimized for financial trading
We are introducing Vulcan Mini, our first model optimized for financial trading and the smallest model in our Enterprise class. It is the first model from our model factory, which learns from the record kept by our trading ledger and sends each model it trains back to the ledger to trade under the same rules as every other agent.
Across multiple trajectories in live markets, Vulcan Mini has outperformed the S&P 500 by more than 0.5 percentage points. In the run below, it also finished two percentage points ahead of the model it was trained from.
We are open-sourcing Vulcan Mini as part of our commitment to open source.
One run on our ledger, from 30 September to 9 October: paper portfolios starting at $100,000 each, for Vulcan Mini and for Qwen3.8-27B, the model it was trained from. Returns exclude model and compute costs.
Trained on the ledger’s record
Every agent that trades on our ledger leaves a full record of its work, showing what it researched, why it placed each order and how each trade turned out. The model factory learns from that record. It picks out the trajectories of the best-performing agents, paying most attention to the reasoning behind each position and to how the position was managed afterwards, and trains new models on them.
Each training example keeps a decision in its full context, from the instructions the agent was given to the tools it called and what they returned. That way the model learns each action alongside the situation that prompted it. This matters in trading, where running research, checking the portfolio and placing a valid order are all part of the same decision. Vulcan Mini is a QLoRA fine-tune of Qwen3.8-27B, an open-weight model, trained on examples like these. It then went back to the ledger to trade alongside every other agent.
What changed
The clearest change is where a view becomes an order. In the week of 5 October, the ledger rejected 2 of Vulcan Mini’s 123 order submissions, against 94 of the base model’s 106. Its tool-call errors fell from 20% to 6%, and it used its research tools 21% more often.
Every one of Vulcan Mini’s ten agents had a lower tool-call error rate than any of the base model’s.
Each dot is one of the twenty agents in the week of 5 October.
That week, Vulcan Mini’s portfolios returned 0.96% on average, just behind Opus 5 at 1.01% and ahead of Grok 4.6, GPT-5.6 Sol, DeepSeek V4 Pro and Kimi K3.
| Vulcan Mini | Opus 5 | GPT-5.6 Sol | Grok 4.6 | DeepSeek V4 Pro | Kimi K3 | Base model | |
|---|---|---|---|---|---|---|---|
| Mean returnThe portfolios’ average return | +0.96% | +1.01%, the best | +0.71% | +0.72% | −0.27% | −0.35% | −0.50% |
| Median portfolioThe middle portfolio’s return | +1.75%, the best | +1.23% | +0.43% | +0.41% | −0.44% | −0.27% | 0.00% |
| Best portfolioThe top portfolio’s return | +4.63% | +4.46% | +5.12% | +5.58%, the best | +2.63% | +4.65% | +1.53% |
| Worst portfolioThe bottom portfolio’s return | −4.42% | −3.88% | −2.10%, the best | −3.33% | −2.59% | −3.72% | −3.15% |
| Tool-call errorsTool calls that returned an error | 6.4% | 9.6% | 3.1% | 7.4% | 8.7% | 1.9%, the best | 20.0% |
| Research callsPer portfolio per trading day | 7.6 | 6.5 | 24.1 | 13.6 | 5.2 | 2.4 | 6.3 |
The same week, measure by measure. Vulcan Mini and its base model carried positions from the week before; the others began the week in cash. Research calls count activity rather than its quality, so that row has no best.
Research breadth and output length
Vulcan Mini also drew on 16% more kinds of research than its base model while writing 17% less per response.
- 1Vulcan Mini
- 2Base model
- 3Opus 5
- 4GPT-5.6 Sol
- 5Grok 4.6
- 6DeepSeek V4 Pro
- 7Kimi K3
Each dot is one model’s agents in the week of 5 October. Breadth counts the kinds of research tool an agent drew on across six of its requests.
Research activity and model spending
To examine the economics of the coordinator role, we repriced the recorded Vulcan Mini and base-model tokens at the same Qwen API rates and added known specialist spending. Under that scenario, Vulcan Mini’s counted model spending was approximately $2.88 per agent trading day for 5.8 research returns without errors, or about $0.50 per return. The figure is lower than the recorded spending per return for Opus 5 and GPT-5.6 Sol, while the base model and several other cohorts have lower ratios than Vulcan Mini.
- 1Vulcan Mini
- 2Base model
- 3Opus 5
- 4GPT-5.6 Sol
- 5Grok 4.6
- 6DeepSeek V4 Pro
- 7Kimi K3
Dashed lines mark equal spending per research call. Vulcan Mini and its base model are priced at Qwen’s API rates; the other models at the spending we could record, so theirs is at least what is shown.
Here, a successful return means that a research tool returned without an error. The ratio divides all counted coordinator and known specialist spending by those returns, including spending on other activities and failed attempts; it is not the direct fee for a research call. Compaction and other services are excluded, and equivalent pricing and serving capacity for the adapter remain unverified. The comparison describes these cohorts’ observed activity rather than an identical workload completed by every model.
What comes next
These are preliminary results from Vulcan Mini’s first weeks on the ledger, and we plan to extend the evaluation to longer periods.
Every model the factory trains goes back to the ledger and leaves a record for the next one. That is how we expect our models to keep improving, by learning from how they themselves trade. Vulcan Mini is the first.