Ant Group releases Ling 3.0 Flash Fin for long-context finance work

Ant Group's InclusionAI lab has released Ling 3.0 Flash Fin, a finance-focused language model built for research, spreadsheet analysis, valuation work, and multi-step agent tasks. The model became available through public API gateways on August 27, with free access advertised through September 25 on Vercel AI Gateway.

Key takeaways

• Ling 3.0 Flash Fin has 124 billion total parameters and activates about 5.1 billion per token.
• Its hosted endpoints expose a 256K-token context window, up to 32K output tokens, reasoning, and function calling.
• Vercel offers free access through September 25, with separate model IDs for continued billing or a hard stop.
• Downloadable weights are promised for the following week but were not available at publication time.

What Ling 3.0 Flash Fin adds

The model keeps the sparse mixture-of-experts architecture used by Ling 3.0 Flash and adds continued training on financial material, domain-specific post-training, and tool-use optimization. InclusionAI positions it for long annual reports, research collections, financial workbooks, information retrieval, investment analysis, valuation modeling, and banking workflows.

Vercel lists a 256K context window and 32K maximum output. OpenRouter reports the exact context limit as 262,144 tokens and confirms support for reasoning and function calls. The model accepts tool definitions, but OpenRouter says it does not enforce structured JSON through response_format.

API access and the free window

Vercel AI Gateway exposes two identifiers with different behavior after September 25. The standard inclusionai/ling-3.0-flash-fin identifier is free during the offer and begins standard billing afterward. The inclusionai/ling-3.0-flash-fin-free identifier is designed to return an error when the promotion ends, preventing an unnoticed switch to paid requests.

OpenRouter also lists a free endpoint, currently served by NovitaAI. Its documentation warns that free endpoints are rate-limited, so teams should not assume promotional access has production-grade capacity or service guarantees.

Financial teams evaluating the model should also separate model capability from data quality. A long context window can hold large reports and workbooks, but it does not guarantee current market data, correct calculations, regulatory suitability, or reliable investment conclusions.

Weights are promised, not released

Chinese technology publication ITHome reports that InclusionAI plans to publish Ling 3.0 Flash Fin's weights during the following week. That is a future lifecycle event, not part of the current API release.

Until the files and their license appear on an official InclusionAI repository or model hub, the model should be described as API-accessible rather than open-weight. The existing Ling 3.0 Flash family does have official Hugging Face artifacts, but those files do not substitute for the finance-tuned checkpoint.

The evidence gap

InclusionAI says it evaluated the model across finance-oriented suites including FinFIRST, FinSearchComp Verified, Finance Agent, APEX-Agents, SpreadsheetBench, and τ³-Banking. Public gateway listings confirm the endpoint, context window, and tool support, but they do not independently verify those quality claims.

No mature third-party benchmark result was available when this article was prepared. That matters for a model aimed at consequential financial work. Early evaluations should use fixed internal tasks, verified source documents, deterministic calculation checks, and human review. Teams should avoid real trades, customer advice, or regulated decisions until the model's accuracy and failure modes are measured against their own controls.

The next confirmation points are the promised weight release, its license and model card, and independent benchmark runs that test financial retrieval and spreadsheet execution rather than general chat quality alone.

Source check

• Vercel's API availability announcement confirms the hosted specifications, model identifiers, and September 25 offer deadline.
• OpenRouter's live model page independently confirms public endpoint availability, context length, tool support, and current free pricing.
• ITHome's launch report corroborates the release and reports InclusionAI's plan to publish weights the following week.