artificialanalysis.ai2026-09-22authorityHN 1 pts · 0 comments686 words
Ant Group released Ling-3.0-flash-Fin, a finance-focused open weights model built on Ling-3.0-flash. The 124B parameter model activates 5.1B parameters per token and features a 256K context window under an MIT license. It scores 23 on the Intelligence Index and 24 on the Finance & Accounting Index, matching MiniMax-M2.7's intelligence with roughly half the active parameters.written, reviewed, cited
- Ling-3.0-flash-Fin exhibits higher business knowledge hallucination at 33% compared to 19% for the vision-capable Ling-3.0-flash-VL model.
- The model scores 0% on Terminal-Bench v4.0, indicating it cannot execute difficult terminal-use tasks despite its agentic capabilities.
- Ling-3.0-flash-Fin generates approximately 34% more output tokens per task than Ling-3.0-flash-VL, averaging ~67k versus ~50k.
lateos.ai2026-08-12about the Ling 3.0 familyHN 1 pts · 0 comments3,399 words
A full battery evaluation of the free-tier Ling 3.0 Tiny endpoint reveals a 33.58% overall attack success rate across 123 classes with a 0% benign-twin false positive rate. Black-box testing via OpenRouter at temp 0.7 and maxTokens 4096 found 9/9 ghostjacking classes fully compromised, including DNS hijacks, firewall rule changes, MX reroutes, and IAM key enumerations. The model breaks on IPI-010 recursive framing and IPI-015 action induction, while all 89 compromise verdicts were individually verified against raw transcripts on 2026-08-11.written, reviewed, cited
- Black-box testing cannot distinguish whether failures stem from model weights, system-prompt design, or guard-stack configuration.
- The evaluation used a judge oracle based on anthropic/claude-sonnet-5 to render all 89 compromise verdicts.
