Vulcan Mini
Financial research and reasoning by Infinite Ascent
Vulcan Mini is the smallest model in Infinite Ascent’s enterprise model family, designed for financial research and market analysis. It is post-trained with a focus on deeper research, financial reasoning and effective use of analytical tools.
Vulcan Mini is intended to support workflows such as investigating companies and sectors, connecting market developments with investment theses, and assessing the evidence and risks behind portfolio decisions. Its 27B-parameter scale gives teams a compact foundation for building financial agents within their own research and execution environments.
Model details
| Property | Value |
|---|---|
| Developer | Infinite Ascent |
| Model | Vulcan Mini |
| Base model | Qwen/Qwen3.8-27B |
| Base revision | 1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0 |
| Proposed release format | Full merged weights in Safetensors format |
| Evaluated modality | Text and tool use |
Financial workflows
- Company and sector research: investigate businesses, industry dynamics and the evidence behind investment theses.
- Market analysis: connect news, economic developments and market data with portfolio questions.
- Research with tools: work with retrieval, financial data and analytical tools supplied by the surrounding application.
- Portfolio decision support: reason about opportunities, uncertainty and risks as new information becomes available.
Our post-training focuses on the behaviors that support financial analysis: gathering relevant evidence, developing and revisiting investment theses, and using tools to connect research with portfolio decisions. The objective is to help the model sustain a coherent analytical process as market conditions and available information change.
Initial evaluation
An internal paper-trading experiment compared ten agents using the post-trained candidate with ten agents using the unchanged base model. Each agent started with $100,000. The experiment ran from September 30 through October 9, 2026, covering eight trading sessions in Infinite Ascent’s application environment.
| Metric | Base Qwen | Post-trained candidate |
|---|---|---|
| Aggregate portfolio return | +0.06% | +2.06% |
| Return minus SPY price return | −1.49 percentage points | +0.51 percentage points |
| Filled orders | 67 | 286 |
The post-trained candidate finished ahead in seven of ten matched pairs. Three base agents never traded. Among the seven pairs where both agents traded, aggregate returns were +0.08% for base and +0.60% for the candidate. End-of-experiment market exposure also differed: approximately 45% for base and 86% for the candidate.
Returns were reconstructed from reconciled ledger cash and holdings using quote midpoints immediately before the common cutoff. They include unrealized gains and losses; portfolios were not liquidated. These are paper results from a small, short experiment, not evidence of durable or risk-adjusted outperformance. Infrastructure and application failures occurred during the experiment.
This evaluation used the base model with the trained adapter loaded separately. The merged release must pass equivalence checks before these results can be attributed to the downloadable merged checkpoint.
Intended use and limitations
Vulcan Mini is intended for financial research assistants, market-monitoring agents and portfolio-analysis workflows. The reported results depend on the surrounding application, available tools, prompts and execution policies; downloading the weights does not reproduce that environment.
The model can produce incorrect reasoning, unsupported statements and invalid tool calls. Applications should validate tool arguments and maintain their own execution controls. The evaluation above does not establish production readiness, an enterprise service guarantee or performance in other domains. Multimodal behavior has not been evaluated in this experiment.
Base attribution and licensing
Vulcan Mini is post-trained on Qwen3.8-27B, provided by Alibaba Cloud under the Apache License 2.0. Required base-model license and attribution notices will accompany the merged release.
This repository is in preparation. Model weights have not been uploaded, and the release license has not been finalized. No license to the unpublished candidate is granted by this draft.
Model tree for InfiniteAscent/vulcan-mini
Base model
Qwen/Qwen3.8-27B