By UNST33R
Which model tier for which security workload
Choosing between unst33r-flash, unst33r-deepseek, and unst33r-pro using Vals AI security benchmarks and your own harness.
All UNST33R tiers sit behind one API key and one OpenAI-compatible endpoint. Every tier supports SSE streaming and tool calling, so switching models is a one-line change to the model field.
The practical goal is picking the lowest-priced available tier that clears the bar for a given task.
The lineup at a glance
| Tier | Base model | Where it fits best |
|---|---|---|
unst33r-flash | GLM-5.3-Flash (uncensored) | Everyday terminal & agent loops: multi-step recon, interactive tool use, drafting and refining payloads. |
unst33r-deepseek | DeepSeek-V4.1-Flash (abliterated) | PoC generation & code-heavy agents: vulnerability reproduction, patch analysis, and fast multi-turn tool execution (up to 384k output). |
unst33r-pro | GLM-5.3 (uncensored) | Deep terminal reasoning: long-horizon shell workflows where smaller models stall or loop. |
See the pricing page for current per-token rates.
What public leaderboards show
Each tier is a reduced-refusal variant of an open-weight model (how abliteration works). While refusal removal changes guardrail behavior, core reasoning and coding capability track the upstream base model closely.
Independent evaluations on Vals AI (as of October 2026) show how the base models split across security and engineering workloads:
| Base model (Tier) | CyberBench v1.1 (PoC & patching) | Terminal-Bench 4.0 (CLI agents) | Vals Index (Overall agentic) |
|---|---|---|---|
GLM-5.3-Flash (unst33r-flash) | 67.86% | 37.37% | 50.86% |
DeepSeek-V4.1-Flash (unst33r-deepseek) | 73.69% (#10 overall) | 31.82% | 61.26% (#1 open-weight) |
GLM-5.3 (unst33r-pro) | 72.08% | 38.89% | 59.73% |
Two takeaways stand out from the public numbers:
- Price doesn't strictly dictate capability on every task. On Vals AI's CyberBench (which grades agents on generating working crash PoCs and source patches for real OSS-Fuzz vulnerabilities),
DeepSeek-V4.1-Flash(73.69%) edges outGLM-5.3(72.08%) in less than half the runtime. Meanwhile, the GLM-5.3 family holds the lead on long-horizon CLI tasks in Terminal-Bench 4.0. - Refusal reduction can affect strict formatting. Public benchmarks grade the unmodified base models. On abliterated and uncensored variants, some models get looser with complex JSON schemas or multi-turn tool calls. Always verify tool-call syntax on your own harness (OpenClaw, Hermes, or custom loops).
Test on your own harness in 10 minutes
Leaderboards tell you where to start; your own prompts tell you what to run in production:
- Pick 5–10 representative tasks from your workflow, including tool definitions if you run agents.
- Run the batch across
unst33r-flash,unst33r-deepseek, andunst33r-proby changing only themodelstring. - Check the console usage table for token counts and exact per-request cost, and pick the cheapest tier that passes both your answer-quality bar and your tool-call formatting checks.
Keep it to authorized scope
These models exist so authorized testers can run a full playbook without fighting mid-run refusals. Review the acceptable-use policy before wiring any tier into an automated harness.
Ready to run a comparison? Get an API key and check the model catalog for copy-paste quickstarts.