Ling 3.0 Flash Fin: Ant Group's Open-Weights Finance Model, Now Independently Scored

Ling 3.0 Flash Fin, the finance-focused mixture-of-experts model from Ant Group's inclu​sionAI lab, now has something it lacked for its first three weeks: a benchmark number from someone other than the lab that built it. Artificial Analysis has published an independent evaluation of Ling 3.0 Flash Fin, scoring it 23 on the Intelligence Index (v4.3) and 24 on the Finance & Accounting Index, and placing it on that index's intelligence-versus-active-parameters Pareto frontier. The model surfaced as an unannounced free API listing on August 27, 2026, and got its MIT-licensed weights on September 3; the outside score is the first real check on whether the finance tuning on top of Ling 3.0 Flash buys anything — and it lands a little differently than the vendor's own framing suggests.

The snapshot

Everything below is drawn from the model card, the Hugging Face repository, and Artificial Analysis's evaluation. The first two are vendor-authored; the third is not.

• What it is — a finance-focused mixture-of-experts model built on the same backbone as Ling 3.0 Flash: 124B total parameters with ~5.1B active per token.

• Where it sits — text-only, a 262,144-token context window, and up to 32,768 output tokens, the same serving shape as its parent.

• What it costs — $0 per 1M input and $0 per 1M output on a one-month free API trial that opened with the August 28 announcement; Artificial Analysis independently lists the same zero price. No paid list price has been published. The BF16 weights are free to download and self-host.

• What it does — reasoning is enabled by default and can be switched off, function calling is supported, and structured JSON output is not enforced.

• What changed — an official launch on August 28 with CICC, MIT-licensed BF16 weights on Hugging Face on September 3 alongside the FinFIRST benchmark, and on September 16 a first independent score from Artificial Analysis.

• What it still lacks — a paid list price, and a third-party run of FinFIRST itself.

What the model card says

The model card lives on Hugging Face under inclu​sionAI's own org, with weights behind it. Its claims are still vendor-authored — Artificial Analysis measures the model and publishes its own index scores, but it does not re-run Ant's FinFIRST suite — so read the card as intent and treat its benchmark numbers as Ant's. It describes Ling 3.0 Flash Fin as the family's "first finance-enhanced model," extending Ling 3.0 Flash "through continued training on high-quality financial data," and centers the release on end-to-end financial research, source-grounded search, multi-document reasoning across filings, and valuation and spreadsheet workflows.

• Architecture — 124B total / ~5.1B active MoE, described as built on Ling 3.0 Flash, which itself uses a hybrid Ki​mi Delta Attention + MLA stack with 1/64 sparse MoE (vendor-described).

• Context — 262,144 tokens in, up to 32,768 tokens out.

• Modality — text in, text out; no image, audio, or video input.

• Reasoning — on by default, can be disabled per request.

• Tool calling — supported, including tools and tool_choice; no response_format support, so piped JSON is not guaranteed.

• Weights and license — MIT; the checkpoint is released in BF16 (about 255 GB across 64 shards) on Hugging Face as inclusionAI/Ling-3.0-flash-Fin, matching the family's permissive open-weights standard.

• Serving — the card says Fin shares Ling 3.0 Flash's architecture and is compatible with the SGLang and vLLM runtimes that already serve the parent. That compatibility is a vendor-stated claim; the speed figure that exists is Artificial Analysis's, not Ant's.

What changed since the release — and what is still unverified

The unknowns that mattered most three weeks ago are closed. Ant Group formally announced Ling 3.0 Flash Fin on August 28 in partnership with CICC, framing it as the family's first finance-enhanced model and opening a one-month free API trial. Then, on September 3, the BF16 checkpoint appeared on inclu​sionAI's Hugging Face org under an MIT license — the same permissive terms as the rest of the Ling 3.0 family — so the model can now be downloaded, self-hosted, and fine-tuned. The missing weights were the single biggest reason to stay cautious; that objection is gone.

The reference for running Fin is the parent shown above. Because the two models share an architecture, the Fin card points to Ling 3.0 Flash's deployment guide rather than publishing a second one, and it states that Fin is compatible with the SGLang and vLLM runtimes that already serve the parent. That is asserted by the vendor, not demonstrated with a Fin-specific production serving benchmark — though Artificial Analysis does report Fin at 162.4 output tokens per second in its own harness, which is faster than the 115 median it tracks for comparable open-weight models.

What makes the release more than a weights drop is FinFIRST — Financial Information Retrieval, Sourcing and Traceability — the benchmark Ant Group built with CICC's investment-banking team and released the same day. It tests whether a model can find the right data, cite authoritative primary sources, run the calculation correctly, and stand behind the process, scoring each answer across 701 granular points that cover result, process, and evidence. The set is small but severe: more than 50 finance professionals wrote and validated the questions, only 9.78% of candidate items survived screening, and the 123 that remain are 61.8% calculation-heavy and 32.5% multi-source. Because the questions, reference answers, and scoring criteria are public under an Apache 2.0 license, FinFIRST gives third parties a shared yardstick to verify Ant's claims with. A tool for independent testing, though, is not a substitute for it — and three weeks on, nobody outside Ant has published a FinFIRST run.

Money is the main open item now. The API is free through the one-month trial, which puts the end of the free window around late September, and no paid list price has been published — so budgeting against Fin past that point still means guessing. Everything else that was open in August has a number attached to it.

Why a finance-tuned Ling 3.0 Flash is a real move

The finance direction is the least surprising thing about this release, because the family was already pointing there. In Ant's own model-card evaluation for Ling 3.0 Flash, the lab reported 27% on τ3-Bench Banking — second among flash-class open models behind DeepSeek V4 Flash — and the highest end-to-end Elo on GDPVal v2-AA at roughly 1,107. Those are vendor-reported numbers on the generalist; a specialist exists to push exactly those numbers up. Ant Group runs AI inside Alipay, payments, risk, and investment products, so a finance-tuned execution model is the family's stated playbook — "planning-execution separation" — applied to the vertical where the company actually makes money.

The screenshot above is how the family's independent reference point looked in August: Artificial Analysis showing Ling 3.0 Flash at 38 on an older build of its Intelligence Index. That figure is now out of date, and the reason is worth understanding before you compare it to anything. Artificial Analysis re-scores models when it revises the index, and its current v4.3 build puts Ling 3.0 Flash at 21 against a median of 8. The 38 is not comparable to a v4.3 number — including Fin's.

What v4.3 says is more interesting than a straight downgrade. Ling 3.0 Flash Fin scores 23 on that index — above its own generalist parent at 21, and just behind Ling 3.0 Flash VL at 25. On the Finance & Accounting Index, Fin and the VL variant tie at 24. Artificial Analysis also places Fin on the Intelligence Index versus active-parameters Pareto frontier: it matches MiniMax-M2.7's score of 23 while activating 5.1B parameters per token against that model's 10B. The nuance cuts the other way too — both Ling models sit below the total-parameter frontier, where Qwen3.8 27B (xhigh) scores 34 with 27B total parameters.

The finance tuning is not a clean sweep, either. Fin's business-knowledge accuracy beats the VL variant's (17% against 11%), but its hallucination rate is worse (33% against 19%), and it trails on the agentic measures: 7% on AutomationBench-AA against 16%, 1,171 Elo against 1,225 on GDPval-AA v2, and 0% on Terminal-Bench v4.0 for both. It also spends roughly 67,000 output tokens per Intelligence Index task, about 34% more than VL. That is the profile of a model pointed hard at document-heavy financial work rather than general agentic execution — which is close to what Ant claims for it, and narrower than "better at finance" sounds.

Free trial plus open weights — what that is worth

A one-month free API trial plus MIT-licensed weights is a materially better position than the free, weightless week this model had at first. The opportunity is twofold: probe the finance tuning on your own multi-step investment tasks at zero token cost on the API, or download the BF16 checkpoint and run — and fine-tune — it yourself. The caution is narrower than it was, but it has not disappeared. One independent evaluation is a real signal, not a production guarantee, and what it shows is a specialization that buys document-grounded accuracy at some cost in hallucination control and agentic execution. Run your own tasks before you route money through it.

This is exactly the situation a routing layer is built for. Through one API that fronts 200+ models, you can put Ling 3.0 Flash Fin on a test path with automatic failover to a proven generalist — when the new model underdelivers or the free tier rate-limits you, traffic shifts without touching application code. When the trial closes and a list price lands, a router that passes provider prices through at 0% markup shows the new number the same day, so your cost model updates without a renegotiation. Try the specialist, keep the safety net, and decide on data instead of vendor copy.

Who should load it up

Finance and agent teams building investment-analysis, research, or long-horizon decision workflows are the obvious audience, and they split into two camps: teams that want a zero-cost hosted trial of the finance tuning before the free window closes, and teams that want to self-host or fine-tune, for whom the open weights are the point. The Artificial Analysis profile sharpens the fit — if your workload is long filings, multi-source reconciliation, and sourced research, Fin's scores are aimed at you; if it is agentic tool use under tight token budgets, the VL variant and the generalist parent both look stronger.

Skip it if you need enforced JSON output (the endpoint does not support response_format), if you need a paid, production-grade API with a known price, or if you are in a regulated context where neither a vendor card nor a single third-party index is sufficient evidence on its own. The open-weights objection that applied in August no longer does — the MIT checkpoint is on Hugging Face. For the most-documented member of the family, Ling 3.0 Flash remains the safer default.

What to watch

The August watchlist was four items long; three are now done. Closed: the official announcement (August 28), the weights release (September 3, MIT), and a first independent evaluation (September 16, Artificial Analysis). Still open, in order of importance: a third-party run of FinFIRST by someone with no stake in Ant, which would test the vendor's own benchmark on neutral ground; a paid list price, which turns the trial math into a budget question; and Fin-specific production serving numbers on vLLM or SGLang, which the card promises by architecture but nobody has published.

The honest status: Ling 3.0 Flash Fin is twenty days old, open-weights under MIT, free on the API for roughly another ten days, and now independently scored at least once. It stopped being a black box in September — the weights are downloadable, FinFIRST hands the industry a shared yardstick, and Artificial Analysis has put a number on the model rather than its backbone. What is real today is that a finance-tuned member of the fastest-moving open-weight family in the market is free to call and free to run, and that the first outside measurement says the specialization is genuine but narrow. The first third-party FinFIRST score and the paid price will turn this from an interesting release into a decision.