Ling-3.0-flash-Fin targets self-hosted financial research

Ant Group’s InclusionAI team has released Ling-3.0-flash-Fin, a finance-tuned version of Ling-3.0-flash. The open-weight model targets source-backed research, multi-filing analysis, valuation workflows, spreadsheet reasoning, and report preparation. Its MIT license permits commercial use, modification, and distribution subject to the license terms.

The release gives developers a finance-specialized model with 124 billion total parameters, 5.1 billion active parameters per token, and a 256K-token context window. That combination can reduce inference compute while preserving enough capacity for long documents. Self-hosting can also keep confidential filings and internal models inside an organization’s chosen environment.

124B weights, 5.1B active

Mixture-of-Experts routing activates a small portion of the network for each token, lowering compute relative to a dense 124B model. All 124B parameters still need storage unless an implementation offloads part of the model. Raw weights require about 248GB at BF16, 124GB at 8-bit precision, or 62GB at 4-bit precision before runtime overhead, KV cache, and long-context memory.

Thinking mode is enabled by default. InclusionAI recommends temperature=1.0, top_p=0.95, and top_k=20 for general inference. Production deployments should retune those settings against representative financial tasks, especially when output length, citation consistency, or deterministic calculations matter.

Built for filings, models, and reports

InclusionAI describes the model as a continued-training variant designed around an analyst’s full research process. Its intended workloads include:

• Source-backed research: retrieving authoritative material and producing traceable answers, supported by the accompanying FinFIRST dataset.
• Multi-document analysis: reconciling reporting periods, accounting definitions, assumptions, and conflicting figures across filings, earnings releases, and annual reports.
• Valuation work: following cross-sheet dependencies, testing scenarios, and explaining changes in assumptions or outputs.
• Report preparation: organizing evidence, calculations, conclusions, and citations into reviewable research materials.

Because the model accepts text only, scanned filings, charts, and spreadsheet files require external tools. A production system will need OCR or document extraction for image-based PDFs, plus an agent or application layer that converts workbook data into model-readable context and writes validated results back to the spreadsheet. Search and citations likewise depend on retrieval infrastructure rather than the model alone.

Benchmarks expose the tradeoffs

Artificial Analysis reports an Intelligence Index score of 23, equal to MiniMax-M2.7 while activating about half as many parameters per token, 5.1B versus 10B. In that benchmark snapshot, Ling-3.0-flash-Fin sits on the intelligence-to-active-parameter Pareto frontier, meaning no tested model delivered a higher index score with the same or fewer active parameters.

Full results and methodology are available on the benchmark page. The finance tuning improved business knowledge accuracy relative to the vision-language sibling, while the lower non-hallucination rate makes citation checks and source verification necessary for factual work.

Output length creates another operational cost. The Intelligence Index evaluation produced 250 million output tokens, compared with a 100 million median across tested models. That works out to roughly 67,000 output tokens per task and about 3.2 times MiniMax-M2.7’s usage. Artificial Analysis measured generation at 162 tokens per second in its environment, though sustained latency and throughput will vary with hardware, quantization, context length, batching, and serving configuration.

The 7% AutomationBench-AA score and 0% Terminal-Bench result also limit its appeal for autonomous workflows spanning SaaS tools, shells, and long sequences of actions. Those scores measure specific agent setups rather than every possible integration, but they support testing the model as a research component before assigning it control of broader toolchains.

Where self-hosting fits

Teams building equity-research search, filing comparison, internal valuation, or report-drafting systems have the clearest reason to evaluate the model. Its low active-parameter count can reduce per-token compute, while local deployment offers control over data location, logging, access policies, and model updates. The large total weight footprint and verbose reasoning still require substantial memory and careful capacity planning.

The MIT license simplifies commercial adoption, although teams should separately review serving software, bundled dependencies, training-data risks, and their own compliance obligations. Financial conclusions, valuation assumptions, and generated calculations require professional review and do not constitute investment advice.

Benchmark it before trusting it

Teams evaluating Ling-3.0-flash-Fin should use a test set drawn from their own filings, models, and research standards. A useful evaluation should:

1. Measure citation accuracy and require page-level evidence or direct quotations.
2. Check period alignment, units, currencies, accounting definitions, and restatements across documents.
3. Run calculations through deterministic code or a spreadsheet engine, then compare them with the model’s reasoning.
4. Test scanned PDFs, long filings, tables, and cross-sheet references through the intended extraction pipeline.
5. Set output and reasoning limits to control latency and token consumption.
6. Profile memory, throughput, and concurrency at realistic context lengths on the target hardware.

Ling-3.0-flash-Fin offers an efficient compute profile for source-backed financial research, provided the deployment can hold its full weights and surround the model with retrieval, document extraction, calculation tools, and human review. Its benchmark results support testing it for analyst-facing workflows, while its hallucination rate, output volume, and weak agent scores argue for constrained use rather than unsupervised automation.