inclusionAI: Ling 3.0 Flash Fin

The FinFIRST benchmark, co-authored by the Ling Team at Inclusion AI (Ant Group), Peking University, and China International Capital Corporation, has been published on arXiv as a companion to the Ling-3.0-flash-Fin model release. FinFIRST (Financial Information Retrieval, Sourcing and Traceability) comprises 123 expert In unified evaluation of 15 model configurations under a shared tool setting, Claude-Opus-5 achieved the highest atomic score (87.59%), while GPT-5.6-Sol was also evaluated (truncated in excerpt). The benchmark is notable for jointly evaluating answers and supporting evidence through atomic criteria, enabling fine-grai

A detailed third-party technical analysis confirms that Ling-3.0-flash-Fin, published by inclusionAI (Ant Group) on 2026-09-03, is a pure continued-training finetune of Ling-3.0-flash with no structural architecture changes. The analysis verifies byte-for-byte identical safetensors metadata, per-dtype parameter breakdo The measured parameter count is 127.49B (across BF16/FP32 dtypes) rather than the announced "124B," and the model card confirms development by Ant Group with leading financial institutions and domain experts. The analysis also introduces FinFIRST, the benchmark inclusionAI open-sourced alongside Ling-3.0-flash-Fin, whi

BenchLM's profile page treats Ling 3.0 Flash Fin as tracked but not publicly ranked, with only three provisional sourced benchmark rows available and no comparative rank assigned as of September 4, 2026. The spec sheet lists a 262K-token context window sourced from Vercel AI Gateway availability, marks the API model ID Coverage is split into eight categories (agentic, coding, reasoning, knowledge, math, multilingual, multimodal, instruction following), of which only agentic has any provisional data while the remaining seven are listed as not measured. The decision snapshot shows the capability field is unranked with a field median of

Artificial Analysis has published an independent evaluation of inclusionAI's 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. According to the report, this is the first outsi The independent AA score provides the first external check on whether the finance-specific continued training on top of Ling 3.0 Flash buys measurable capability, and it lands slightly differently than the vendor's own framing suggests. The model is described as a mixture-of-experts finance-tuned variant from Ant Group

Vercel's changelog, as recapped by CreateWith, lists Ling 3.0 Flash Fin from Inclusion AI as available through Vercel AI Gateway and free to use through September 25, 2026. The model combines a 256K-token context window with reasoning, function calling, and up to 32K tokens of output, specifications that Vercel positio Developers integrating through Vercel must pick between two model identifiers, and the free window lets small teams prototype document-review assistants, internal research tools, and financial-data agents before committing to paid usage. The recap cautions that any finance model output should still be checked against o

Model Pulse tracks InclusionAI's Ling 3.0 Flash Fin across multiple providers and confirms its availability through Kilo Gateway with a listed price of $0.06 per million input tokens and $0.18 per million output tokens, plus a $0.012 cache-read rate. The Kilo Gateway variant exposes a 262K-token context window with rea According to the recent changes log, the Ling 3.0 Flash Fin listing was first surfaced via Kilo on September 3, 2026, with an earlier provider entry dated August 27, 2026, and a Vercel follow-up on August 28, 2026. Best published input and output price history shows the lowest listed price holding flat at $0.06/$0.18 p

An LLM Explorer model-card listing for inclusionAI/Ling-3.0-flash-Fin reports the model as a 127.5B-parameter BailingMoeV3 hybrid MoE finetune of inclusionAI/Ling-3.0-flash, with 186 GB required VRAM, 262,144-token context length, MIT license, and Transformers 4.45.0 compatibility — confirming the architecture, license The page enumerates the 64 safetensors shards (sizes ranging 3.9–5.3 GB), a vocabulary of 157,184, and bfloat16 torch data type, and lists sibling Ling variants (Ling 3.0 Flash base/midtrain/Fin int4/Fin Fp8/CIRU/Heretic) for context on quantization and specialization paths. Tags include Agents, Bailing, hybrid, Financ