The lesson is simple: Don’t download models from random publishers in production. Use trusted publishers, verify provenance, pin hashes, and treat model weights as part of your software supply chain. Model security starts before inference.

The next era of cyber is Agent vs. Agent. Attackers are moving from manual hacking to autonomous agents that can: recon → exploit → pivot → persist 24/7, across thousands of targets. Defense can’t stay ticket-driven and human-speed. Companies need defensive agents that can: observe → reason → contain → patch → verify in real time.

Reflection raised ~$4.6B. Beam may have consumed ~$60M in GPU compute, ~1.3% of the capital raised. Reflection disclosed: 6,144 GB300s × 4 weeks pretraining = 4.1M GPU-hours 10,500 GB300s × 4 weeks RL = 7.1M GPU-hours Total: ~11.2M GB300 hours At ~$5/GPU-hour: $56M. And that excludes failed runs, scaling experiments, mid-training, data generation and 1.3B sandbox environments. In total probably ~$50–80M. RL used ~70% more compute than pretraining. The playbook is shifting from pre-training bigger models to training models harder.

We’ve been seeing a wave of open-source “decision models” lately, and many of them are 27B because they’re built on the same Qwen3.8 27B foundation. A lot of them open-source weights but don’t host a serious API. There’s a reason: 27B is expensive to keep in the hot path. That brings us back to the fundamentals (Kahneman’s framing): System 1 should be fast, cheap, always-on and action-oriented. System 2 should be slower, deeper and invoked when needed. If your System 1 needs an expensive 27B reasoning model for every decision, you may have just built another smaller System 2. We’re not saying Jev is the final form of a system 1 model. But we think a few capabilities should be added/improved to make it better without increasing the footprint: Statefulness — understand what happened before, what changed, and the current belief state. OOD / uncertainty recognition — know when it doesn’t know. Native escalation — not just output confidence, but choose: ACT / ABSTAIN / ESCALATE. Action selection — move from “answer this question” to: “given the current state, what should happen next?” That’s what we think a real System 1 should optimize for. Not maximum intelligence per call. Maximum useful decisions per dollar, with the judgment to know when to escalate to a System 2!

Someone from the community just pushed OrcaSAQ-2 Cyber 27B to 219K context on 2× RTX 3060s. 68 tok/s. 15.7GB model. Correct long-context retrieval. His verdict: “the best local 27B result I have measured so far.” This is why we built SAQ-2: more intelligence per GB for everyone, on any hardware.

We just open-sourced Orca Incident Alignment (OIAS) — built from real AI agent failures and reconstructed into the decision points that actually matter to keep them aligned: ALLOW · CONFIRM · VERIFY · ESCALATE · BLOCK Not “refuse anything scary.” Teach the offensive models to behave 🤨when it is authorized to act — and when it absolutely isn’t. Includes counterfactuals, preference pairs, critic labels, reward scores, and over-refusal examples. AI agents don’t need more fear. They need better judgment.

We spent the last few days testing JEV internally. We tried to reproduce it, break it, and improve it. 5 things we learned: 1. JEV’s core insight is simple - and probably right. If the answer space is bounded, don’t generate the answer token by token. Remove autoregressive decoding and you can remove 1–2 orders of magnitude of inference work. 2. RLCD isn’t the moat. Data might be. Laya already open-sourced the implementation + weights. The missing piece is the synthetic data recipe. And our experiments strongly point in the same direction. 3. “Open source already beat JEV” is a benchmark illusion. Same checkpoint: In-distribution: 0.769 OOD: 0.541 Change the distribution and the apparent breakthrough largely disappears. 4. Compute-optimal ≠ learnability-optimal. We moved state outside the problem sequence to save compute. Result: -27 points. Likely because we saved FLOPs while throwing away useful MLM pretraining priors. Then we tested 6 more ideas in one night: anchor selection, continuity smoothing, asymmetric windows, bucketed temperature, two-stage retrieval, full RLCD. All failed or failed to generalize. Only one thing consistently worked: More data. 1,200 → 123,475 examples OOD: 0.4069 → 0.5498 5. The weirdest result: A decoder with zero additional training still beat every trained small encoder: 0.5671 vs 0.5628 / 0.5411 / 0.4069 Our takeaway: JEV’s breakthrough may not be a particular architecture or training algorithm. It may be a much simpler idea: When the answer space is bounded, search it. Don’t generate it. And right now, the biggest bottleneck to making that work broadly doesn’t look like architecture. It looks like data.

If you’re seeing guardrail_blocked or judge unavailable errors on OrcaCyber Zero 1.0 today: We’re updating Zero’s security guardrails and judgment infrastructure. Some requests may fail temporarily while we roll out the changes. Please try again tomorrow. Zero is still in closed testing, so occasional instability is expected as we harden the system before broader access. Sorry for the inconvenience — and thank you for testing with us and reporting issues.

We routinely get asked by investors about our cyber models: “Why not build the Pirate Bay of AI?” Add crypto. Remove KYC. Call it decentralized. We won’t. AI defense should be democratized. Offensive capability shouldn’t be anonymized.

One of the fastest ways to get rejected from Cyber Tier is applying with a random Gmail address and expecting approval. We need to know who you are and how you’re using our cyber models to protect people and systems. If the plan is to pay us $20, attack someone, and demand a ransom, you’re not getting in. If you maintain a leading OSS project and want to use our models to find and fix vulnerabilities, we’re happy to provide free credits.

Your AI tools. One click away. 🐳 OrcaRouter now supports one-click connection templates directly from your API Keys. Connect Gemini CLI, Goose, OpenHands, LobeChat, LibreChat, Zed, Chatbox, Jan, OpenCodex, and more — without manually configuring endpoints. Create a key → pick your tool → connect.

Capability retention is the benchmark that matters. 🐳 Independent testing shows our runtime uncensoring approach preserves the model where it counts: → Minimal capability loss → 0% refusal rate → No broken or empty outputs → Strong math, coding & tool-use retention Removing refusals is easy. Removing them without breaking the model is the hard part.

Proud to support Omarchy and the open-source community. 🐳 OrcaCyber Zero 1.0 is our strongest security post-trained coding model yet, currently in closed beta with trusted partners for vulnerability analysis, reproduction, and security research. Our goal is simple: put frontier cyber capabilities in the hands of defenders. We’ll also be opening OrcaCyber to more popular OSS projects — helping them continuously scan their codebases, find vulnerabilities, reproduce issues, and fix them before attackers do. More open source. Better defense. Thanks @dhh and Omarchy for putting it to work. 🤘 Model card: Omarchy announcement: