How to Run Ling 3.0 Flash Fin Locally

Built from InclusionAI's original weights with our own importance matrix. The calibration corpora behind our builds are public.

• Ling 3.0 Flash Fin is InclusionAI's finance-enhanced Ling model for research, valuation, spreadsheets, and long-horizon agent workflows.
• These GGUFs are self-quantized from InclusionAI's original BF16 weights with our own importance matrix.
• The repo currently includes BF16, AD-Q8_0, and AD-Q6_K builds; lower-bit quants are still uploading.

---

Ling 3.0 Flash Fin, self-quantized to GGUF by Atomic Chat. It has 124B total parameters, activates 5.1B per token, and supports a 256K context window. The checkpoint extends Ling 3.0 Flash with continued training on high-quality financial data.

Highlights

• End-to-end financial research: connects retrieval, evidence review, calculation, modeling, and report preparation in one workflow.
• Source-grounded search: prioritizes authoritative sources and traceable answers. InclusionAI publishes FinFIRST for transparent evaluation.
• Multi-document reasoning: reconciles periods, definitions, assumptions, and conflicting figures across filings, earnings materials, and research.
• Valuation and spreadsheet workflows: understands formulas, estimate updates, cross-sheet dependencies, balance checks, and scenario analysis.
• Research-ready output: separates facts, analysis, judgments, and charts into material that can be reviewed and edited.

[!NOTE] These GGUFs are self-quantized from the original weights, not a repack.

[!IMPORTANT] Thinking mode is enabled by default. InclusionAI recommends temperature=1.0, top_p=0.95, and top_k=20.

Model overview

Pick a file

The download must fit alongside the context cache and runtime overhead. Leave several gigabytes of headroom beyond the file size.

Get started

Run Ling 3.0 Flash Fin locally with:

• Atomic Chat: open the app, search AtomicChat/Ling-3.0-flash-Fin-GGUF, pick a build, and select Use this model.
• llama.cpp: download the AD-Q6_K folder and open its first shard with llama-server.
• LM Studio / Jan: search the repo ID and download the build that fits your machine.

Download the recommended build:

hf download AtomicChat/Ling-3.0-flash-Fin-GGUF \
  --include "AD-Q6_K/*" \
  --local-dir Ling-3.0-flash-Fin-GGUF

Run it:

llama-server \
  -m Ling-3.0-flash-Fin-GGUF/AD-Q6_K/Ling-3.0-flash-Fin-AD-Q6_K-00001-of-00003.gguf \
  --jinja -ngl 99 -c 32768

The chat template is embedded in the GGUF. Keep --jinja enabled so thinking mode and tool-call formatting are applied correctly.

Best practices

Allocate enough output length for research and agent tasks. Financial conclusions, valuation assumptions, and investment decisions still require professional review.

What the model is built for

The producer evaluates the model on FinFIRST, FinSearchComp Verified, FinCRAFT, Finance Agent, APEX-Agents, SpreadsheetBench, and tau3-Banking. These benchmarks cover source-grounded retrieval, investment research, long-horizon execution, valuation modeling, spreadsheet operations, and banking workflows.

See the official model card for the producer's benchmark results, deployment guidance, and limitations.

How these were made

1. Start from inclusionAI/Ling-3.0-flash-Fin, the original BF16 checkpoint.
2. Convert the checkpoint directly to GGUF.
3. Build a per-tensor importance matrix over the public Atomic Chat calibration corpora.
4. Quantize the shipped builds and validate them against the BF16 reference.

The raw evaluation logs currently available for the uploaded builds are included in the logs/ directory of this repository.

License

Released by InclusionAI under the MIT License. Quantized by Atomic Chat.