Zro: Uncensored Security Models

About Zro: Uncensored Security Models

Zro is an inference platform built for coding agents and security teams, with a stated policy of zero request retention. It runs open models across multiple regions, so prompts and code aren't stored. The latest launch introduces Dolly 1 Security and Dolly 2 Security, two abliterated variants of GLM-5.3 Flash and GLM-5.3 designed for red-team and defensive testing workflows where general-purpose models typically refuse requests.

Review

Zro positions its service around speed and data handling rather than model fine-tuning. The team has shipped over 250 billion tokens processed since their July launch, and they've contributed pull requests upstream to inference engines like SGLang and vLLM. This security-focused expansion addresses a specific gap: running uncensored models through an API that doesn't log what you send.

Key Features

• Two abliterated security models - Dolly 1 Security (based on GLM-5.3 Flash) and Dolly 2 Security (based on GLM-5.3, abliterated in-house) - that won't refuse red-team or defensive testing prompts.
• Zero request retention policy across all inference calls, with multi-region execution available.
• Published model weights and evaluations for Dolly 2 Security, hosted on Hugging Face in FP8 format.
• CLI tool that launches coding agents like Claude Code with a single command, pointing it at a Dolly model.
• Free daily usage tier on Flash-based models, with a referral program that credits both parties $5.

Pricing and Value

Dolly 1 Security runs at $0.45 per million input tokens and $1.50 per million output tokens. Dolly 2 Security costs $2.10 input and $6.30 output per million tokens. A free plan covers daily usage on Flash models. Paid Pro and Max plans exist, and the Product Hunt launch offers 25% off the first month with code ZROPRODUCTHUNT. The referral program gives $5 in usage credit to both the referrer and the new user when the latter starts a paid subscription.

Pros

• Clear, per-token pricing for both security models with no hidden retention of prompts or code.
• Dolly 2's weights and evaluation are public, letting teams verify what the abliteration process changed.
• The CLI integrates directly with agent workflows - one command swaps the model under Claude Code.
• Team publishes PRs to upstream inference projects, which gives some visibility into their stack decisions.
• Free tier and referral credits lower the barrier to testing the models before committing to paid usage.

Cons

• Evaluations comparing security-task accuracy against ordinary coding accuracy side-by-side aren't published yet, which makes it hard to spot capability regressions from abliteration.
• The roadmap includes an uncensored DeepSeek-V4.1-Flash and Kimi K3.1, but neither is available now - the current security lineup is limited to the two GLM-5.3 variants.
• This tool is not well suited for teams that require on-premise or air-gapped deployment, since everything runs through Zro's cloud API and CLI.

Zro's security models fill a narrow but real need for security researchers who hit refusal walls with standard APIs. The published weights and zero-retention policy address two common concerns in red-team work, though the evaluation gap means teams will need to run their own benchmarks to confirm Dolly 2 doesn't lose too much general capability. For anyone already using coding agents and running into model refusals during security tests, the free tier makes it practical to try without upfront cost.