Model reference · open weights
Ling-3.0-flash-Fin is an open-weight language model from inclusionAI. Ling-3.0-flash-Fin (BF16) weighs 255 GB; the smallest configuration that runs it is 2× H200 141 GB.
For self-hosting, compare this build's requirements below with the GPU you plan to rent.
Summary of the inclusionAI/Ling-3.0-flash-Fin-int4 model card, 2026-10-01. The estimate below is for another build of the family.
What it is
| Released by | inclusionAI |
|---|---|
| Released | 2026-09-09 |
| Parameters | 127.5B |
| VRAM | 255 GB for the weights |
What it runs on
| Card | Requests at once | Context max | Memory | |
|---|---|---|---|---|
| 8K each | 32K each | |||
| RTX 3060 12 GB … B200 180 GB 12 smaller cards | ||||
| 2× H200 141 GB tensor parallel | 8 | 2 | 68K | 138 GB a card |
| 4× H100 80 GB tensor parallel | 18 | 4 | 146K | 78.1 GB a card |
| 4× A100 80 GB tensor parallel | 33 | 8 | all 256K | 78.2 GB a card |
| 2× B200 180 GB tensor parallel | 98 | 24 | all 256K | 176 GB a card |
| 4× RTX PRO 6000 Blackwell 96 GB tensor parallel | 57 | 14 | all 256K | 93.8 GB a card |
| Requests at once | 8K tokens each | 32K tokens each |
|---|---|---|
| 1 | 257 GB | 258 GB |
| 5 | 258 GB | 264 GB |
| 8 | 259 GB | 269 GB |
| 16 | 263 GB | 282 GB |
| 32 | 269 GB | 307 GB |
| 64 | 282 GB | 358 GB |
One card, with vLLM's small-card settings.
Builds
| Build | Params | Precision | Weights | Smallest setup |
|---|---|---|---|---|
| Ling-3.0-flash-Fin-int4 ↗ | 127.5B | INT4 | 77.0 GB | 2× L40S 48 GB |
| Ling-3.0-flash-Fin-fp8 ↗ | 127.5B | FP8 | 128 GB | H200 141 GB |
| Ling-3.0-flash-Fin (above) ↗ | 127.5B | BF16 | 255 GB | 2× H200 141 GB |
From the model card
Ling-3.0-flash-Fin is the first finance-enhanced model in the Ant Ling family. Developed by Ant Group with leading financial institutions and domain experts, it extends Ling-3.0-flash through continued training on high-quality financial data.
With 124B total parameters, 5.1B activated parameters, and a 256K context window, the model combines financial expertise with efficient inference for long-horizon agent workflows.
Ling-3.0-flash-Fin was evaluated across FinFIRST, FinSearchComp Verified, FinCRAFT, Finance Agent, APEX-Agents, SpreadsheetBench, and τ³-Banking. These benchmarks cover source-grounded retrieval, investment research, long-horizon execution, valuation modeling, spreadsheet operations, and banking workflows. The model is competitive with both similarly sized models and substantially larger general-purpose models, with particular strength in source selection and tool-intensive financial tasks.
The current checkpoint is released in BF16. Because Ling-3.0-flash-Fin shares the same architecture as Ling-3.0-flash, it is compatible with the same SGLang and vLLM runtimes. For deployment instructions, see the Ling-3.0-flash deployment guide.
Important: Thinking mode is enabled by default. For optimal performance, we strongly recommend using
temperature=1.0,top_p=0.95, andtop_k=20for general inference.
We evaluate the quantized models using several datasets. The FP8 quantized model is applied via the blockwise quantization, and INT4 and FP4 models are applied via groupwise quantization with routed experts weights.
| dataset | BF16 | FP8 | INT4 | FP4 |
|---|---|---|---|---|
| GPQA-diamond | 86.30 | 85.26 | 85.39 | 84.38 |
| SciCode | 41.84 | 42.70 | 41.41 | 41.24 |
| FinCRAFT | 54.23 | 55.91 | 54.48 | 54.73 |
| FSC-verified | 78.28 | 77.48 | 76.86 | 76.77 |
As our first finance-enhanced release, Ling-3.0-flash-Fin still requires further validation in complex, long-horizon workflows. Key assumptions, valuation results, and investment conclusions require professional review and do not constitute investment advice.
Future releases will explore finance-enhanced models at larger scales to further improve complex reasoning and long-horizon task execution.
Quoted from the model card on Hugging Face. The full card is behind the Hugging Face link above.