Introducing OrcaC2 1.0, our cyber decision model trained for blockchain security, MDR, incident response, security reasoning, and security operations. Vision support. Up to a 1M-token context window. Native temporal awareness to connect events as attacks unfold (complex reasoning across event sequences). Evaluated on 17 benchmarks spanning cybersecurity, general reasoning,... and blockchain security. OrcaC2 leads the tested baselines across all reported tasks, achieving 90.6 on vulnerability severity assessment and 87.3% overall blockchain security accuracy. AI attacks evolve in real time. Autonomous defense swarms need a radar that can track that evolution, reason across events, and guide the response. We’re building OrcaC2 to be that foundation. Try the API:show more

OrcaC2 1.0 benchmark results: 17 benchmarks spanning cybersecurity, general reasoning, and blockchain security.

OrcaC2 1.0 is available exclusively through OrcaRouter, with zero data retention (ZDR) enabled by default. Start building:

every security ops is a long-context problem: weeks of logs, one needle. relevant launch today: O(1) per-token inference, 100% single-needle retrieval at 1B tokens

Any chance this will be open sourced in the future?

Larp ?

Fix your website!!! Cannot apply to cyber

认同罗翔老师的观点

实战出真知

这种模型别只看参数，得先看它怎么处理长周期数据//纯手搓，关必回。

Congrats on the launch. If OrcaC2 is served behind an OpenAI-compatible API, it can drive Frame, the open-source security agent that's state of the art on BAC-Bench (22 of 30 real access-control bugs; CodeQL and Semgrep find 0).

If I have to pick up the joystick, it isn’t gonna be beautiful. We’ll just say that much.

Bro.. I would hang my IT director if he even thought of presenting this vibe coded bullshit to me.

dumb question, but what am i looking at. shows the number 1 model is DeepSeek. Guessing this leaderboard has nothing to do with security? I see models like "OrcaCyber: Hy4 " but what even is that, like how large is it, what was it trained on? We have Mythos at home.

This leaderboard has nothing to do with security and is more general purpose. OrcaCyber models are all security specialist models which either are post trained (Zero, Ultra) or abliterated. The leaderboard tracks internal and external data for a general intelligence composite score of each model.

Which Jev and Clef versions were used here? I'd also be interested in how temporal awareness was evaluated, especially whether changing event order changes the verdicts.

1M上下文加时间感知 这配置有点东西

上下文能装一百万 我连昨天密码都想不起来

Impressive work with OrcaC2 1.0! The focus on blockchain security is key. 🌐🔒

🤪

100万token上下文，这配置有点猛啊。

这点我认同：介绍 OrcaC2

比起照片元数据，背景里的店招和路牌可能更容易暴露位置😂

百万级上下文，这本地推理能力有点强啊。

这上下文窗口真大，处理安全事件应该挺猛。