FUMEA-F: Frontier Unified Multi-Expert Agent — Financial

FUMEA-F is a 4-expert Mixture-of-Experts language model designed for financial analysis, marketing intelligence, market trend detection, and multi-step financial reasoning. It uses top-2 expert routing per token, activating roughly half its total parameters per forward pass.

Architecture

Property	Value
Architecture	Decoder-only Transformer (MoE)
Number of Experts	4
Experts per Token	2 (top-2 routing)
Gate Mode	Hidden state routing
Context Window	131,072 tokens
Positional Encoding	RoPE with YaRN scaling (factor 4.0, base 32,768)
Precision	bfloat16

Expert Domains

Expert	Domain
0	Marketing intelligence, brand strategy, campaign analysis
1	Financial forecasting, time-series, technical indicators
2	E-commerce trends, consumer behavior, competitive pricing
3	Financial reasoning, chain-of-thought valuation, compliance

Tool Use

Supports structured function calling via chat template with <tool_call> tags. Built-in schemas: analyze_ohlcv, web_search, code_executor, file_reader.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model = AutoModelForCausalLM.from_pretrained(
    "uaytug/fumea-f", torch_dtype=torch.bfloat16,
    device_map="auto", trust_remote_code=True,
)
tokenizer = AutoTokenizer.from_pretrained("uaytug/fumea-f", trust_remote_code=True)

messages = [{"role": "user", "content": "Walk me through a DCF valuation for a SaaS company."}]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
output = model.generate(**inputs, max_new_tokens=2048, temperature=0.6, top_p=0.9)
print(tokenizer.decode(output[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True))

Quantized Versions

See uaytug/fumea-f-gguf for GGUF quantizations from F16 to IQ1_S.

Generation Defaults

Parameter	Value
temperature	0.6
top_p	0.9
repetition_penalty	1.1
max_new_tokens	8192

Limitations

• Merged model, not fine-tuned on curated datasets. Quality depends on source expert capabilities.
• Financial output should not be used as sole basis for investment decisions.
• Extended context performance not systematically benchmarked.

License

Apache 2.0