BLUM Finance 4B

Financial Reasoning, Risk & Trading Intelligence

BLUM Finance 4B is an open financial reasoning model adapted from mlx-community/Qwen3-4B-4bit. It turns timestamped evidence into a balanced thesis, contradictory evidence, explicit risks, invalidation conditions and an explicit confidence value.

It is not a price oracle, broker, autonomous execution system or source of live market truth. BLUM Engine validates market facts and calculations.

Quick Start

pip install "mlx-lm>=0.31,<0.32"

from mlx_lm import load, generate

model, tokenizer = load("Italianhype/Blum")
prompt = tokenizer.apply_chat_template(
    [
        {
            "role": "system",
            "content": "Use only supplied point-in-time evidence. Return JSON.",
        },
        {
            "role": "user",
            "content": '{"ticker":"NVDA","as_of":"2026-07-28T10:00:00Z","evidence":[{"type":"risk","value":"Volume confirmation is missing."}]}',
        },
    ],
    tokenize=False,
    add_generation_prompt=True,
    enable_thinking=False,
)
print(generate(model, tokenizer, prompt=prompt, max_tokens=512, verbose=False))

Measured Evaluation

The candidate was compared with the exact base revision 4dcb3d101c2a062e5c1d4bb173588c54ea6c4d25 on an untouched temporal test split.

Metric	Value
Test samples	53
Base aggregate	0.0
BLUM aggregate	0.962601
Delta	0.962601
No-fabrication	0.928571
Structured validity	1.0
Calibration error	Not measured (0 labeled outcomes)

Evaluation traces

The base score can be low when otherwise readable prose fails the required JSON contract. The aggregate delta therefore measures BLUM task adherence, evidence handling and safety structure on this test; it is not a general intelligence comparison with the base model.

These language-model results measure financial reasoning behavior. This evaluation does not prove trading alpha, future profitability or benchmark outperformance.

Intended Use

• financial thesis drafting from supplied evidence;
• bull/bear contradiction analysis;
• risk and technical invalidation explanation;
• portfolio and benchmark-aware research;
• conservative abstention when evidence is incomplete;
• English and Italian financial analysis.

Limitations

• The model can still produce incorrect or stale statements.

• It has no guaranteed access to current prices, news or filings.

• Numerical levels must be validated by the calling application.

• The current held-out set has no labeled market outcomes, so confidence is conservatively capped but not empirically calibrated.

• It must not independently authorize or execute a financial transaction.

• Performance can vary by market, regime, language and prompt structure.

• This release uses the MLX runtime and requires Apple Silicon. A portable Transformers/GGUF release remains a separate validation target.

Known Failure Cases

• Missing or contradictory point-in-time evidence can force an abstention.
• Unsupported prompts may still elicit plausible but incorrect market facts.
• Long filings can exceed the context supplied by the caller.
• The model is not validated for broker execution or unattended capital control.

Related BLUM Resources

• BLUM Engine Space
• Reasoning dataset

Community Memory

Inference and installation send no telemetry. Users may explicitly create a redacted contribution with blum-contribute. Contributions enter Italianhype/Blum-Finance-Memory in quarantine and never update active weights without review, versioning and out-of-sample validation.

Incremental learning is explicit opt-in. Accepted, deduplicated contributions can join a future immutable training snapshot only after provenance, privacy, schema and quality checks. Each training cycle creates a versioned challenger, compares it with the active model on an untouched test split and promotes it only when the release gates pass. A failed challenger is retained as evidence and cannot replace the active model.

BLUM never trains inside an inference request or lets anonymous input mutate released weights.

Lineage

• Base revision: 4dcb3d101c2a062e5c1d4bb173588c54ea6c4d25
• Candidate revision: ea297ba88ab008e97104b0c118103eef2f8f9ec1
• Dataset revision: 76ad77699d498fc930daf02e452fe3ec8b490f90
• Dataset: Italianhype/Blum-Finance-Reasoning

Disclaimer

For research and educational use. This model does not provide financial advice, guaranteed signals or guaranteed profits.

Citation

@misc{blum_finance_4b,
  title = {BLUM Finance 4B},
  author = {Italianhype},
  year = {2026},
  url = {https://huggingface.co/Italianhype/Blum}
}