A QLoRA fine-tune of Qwen/Qwen3-30B-A3B specialized for financial statement analysis and equity-research-style reasoning. It produces structured, analyst-grade commentary on earnings, fundamentals, SEC filings (MD&A, risk factors), and company comparisons.

Scope note: This is a narrow domain specialist for direct financial-analysis Q&A. It is not an agentic/tool-using model and should not be used for real-time market data, retrieval-augmented workflows, or autonomous agents (see Limitations).

Model details


Base model	Qwen/Qwen3-30B-A3B (MoE, 30.5B total / ~3B active params, 128 experts)
Method	QLoRA (4-bit NF4), attention-only LoRA
LoRA target modules	`q_proj`, `k_proj`, `v_proj`, `o_proj`
LoRA config	rank 32, alpha 64, dropout 0.05, rsLoRA
Trainable params	~27M (0.09% of base)
Epochs / examples	3 epochs / 995 train (53 val)
Final loss	train 0.67, eval 0.67 (closely tracked — no overfitting)
Formats	LoRA adapter, GGUF `q4_k_m`, GGUF `q3_k_m`

Files / quant guide

File	Size	Fits	Notes
`gguf/qwen3-30b-finance-q4_k_m.gguf`	~18 GB	24 GB+ VRAM (e.g. RTX 4090/5090)	best quality
`gguf/qwen3-30b-finance-q3_k_m.gguf`	~14 GB	16 GB VRAM (e.g. RTX 5080) fully on GPU	small quality drop
`adapter_model.safetensors`	103 MB	apply to `Qwen/Qwen3-30B-A3B`	for Transformers/PEFT

On a 16 GB card use q3_k_m for full-GPU speed. The q4_k_m exceeds 16 GB and would need partial CPU offload (--n-gpu-layers below max) — workable on this MoE but slower.

Intended use

Direct, single-turn financial-analysis instruction following, e.g.:

Assess an earnings beat/miss given EPS actual vs estimate
Summarize a 10-K/10-Q MD&A excerpt (revenue/margin drivers, outlook)
Identify and rank material risks from a risk-factors disclosure
Evaluate a company's fundamentals (valuation, profitability, growth, health)
Compare same-sector companies on relative strength

Prompt format (Alpaca-style)

The model was trained on the Alpaca instruction template and follows it reliably:

Below is an instruction that describes a financial analysis task. Write a response that appropriately completes the request.

### Instruction:
{instruction}

### Input:
{input}

### Response:

How to use

Option A — GGUF with llama.cpp

bash

./llama-cli -m gguf/qwen3-30b-finance-q4_k_m.gguf -n 512 --temp 0 -p "$(cat <<'EOF'
Below is an instruction that describes a financial analysis task. Write a response that appropriately completes the request.

### Instruction:
Assess the quality and implications of the company's earnings.

### Input:
Company: Example Corp (EXMP)
Reported EPS: $2.10
Estimated EPS: $1.85
Surprise: +13.5%

### Response:
EOF
)"

Option B — LoRA adapter with Transformers + PEFT

python

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

REPO = "joexie/Qwen3-30B-A3B-Finance"
model = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen3-30B-A3B", torch_dtype=torch.bfloat16, device_map="auto"
)
model = PeftModel.from_pretrained(model, REPO)
tok = AutoTokenizer.from_pretrained(REPO)

prompt = """Below is an instruction that describes a financial analysis task. Write a response that appropriately completes the request.

### Instruction:
Identify and prioritize the most material risks from this disclosure.

### Input:
<paste a risk-factors excerpt or financial data here>

### Response:
"""
inputs = tok(prompt, return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=512, do_sample=False)
print(tok.decode(out[0][inputs.input_ids.shape[1]:], skip_special_tokens=True))

Always put the data to analyze in the ### Input: block. The model reasons over what you give it; it has no retrieval and will fabricate figures if asked for current market state.

Example output

Instruction: Based on the following MD&A excerpt from Honeywell's 10-Q (Q1 2026), identify the key business developments, financial trends, and management outlook.

Input (excerpt): "Net sales increased due to increased pricing... favorable foreign currency translation (Australian dollar, Chinese renminbi, Canadian dollar)... partially offset by lower sales from the divestiture of the PPE business..."

Model output (abridged):

1. Revenue & Margin Drivers

Positive Drivers:

Pricing Power: Increased pricing and price adjustments to offset inflation — the ability to pass cost increases to customers.

FX Favorability: A weakening U.S. dollar boosted international sales (AUD, CNY, CAD strength).

Negative Headwinds:

Volume Declines: Lower sales volumes dragged on growth, suggesting softer underlying demand.

Divestiture Impact: The PPE business divestiture structurally reduced the revenue base.

The model reliably returns this kind of structured, source-grounded analysis in the instruction format. Its measured advantage over the base is consistency and format adherence (it answered 10/10 held-out prompts vs the base's 4/10), not a dramatic gap in analytical depth — see EVALUATION.md.

Training data

~1,048 instruction/response pairs (post-filter) built from public sources:

Yahoo Finance — fundamentals, quarterly earnings (EPS actual vs estimate), price history
SEC EDGAR — 10-K / 10-Q filings; MD&A, risk-factors, and business sections extracted via a prose-scoring parser (skips tables-of-contents and boilerplate)

Composition: earnings 477, MD&A 268, risk 206, fundamentals 60, comparative 18, news 19. Analysis targets for filing-based tasks were synthetically generated and then filtered to remove refusals/meta-commentary.

Limitations and out-of-scope use

Not an agent / no tool use. Training contained zero tool-call examples. The model does not reliably perform web search, function calling, or multi-step agentic workflows. In agentic harnesses it has been observed to hallucinate tool arguments and confabulate data. Do not use it as the backend for autonomous agents, chat assistants with tools, or scheduled report generators.
No real-time data / will hallucinate market figures. It has no retrieval and will fabricate confident-sounding prices, percentages, and headlines if asked for current market state. Always supply the data to analyze in the prompt.
Not investment advice. Outputs are illustrative analysis, may contain errors, and must not be used for trading or financial decisions.
Inherits base limitations — knowledge cutoff and biases of Qwen3-30B-A3B.
English only, tested on US-listed large caps.

License

Apache 2.0, inheriting the base model's license.

xerus19573/Qwen3-30B-A3B-Finance

Qwen3-30B-A3B-Finance