unst33r-deepseek is live: Abliterated DeepSeek V4.1 Flash for security agents

unst33r-deepseek is now available on the UNST33R API. It serves an abliterated variant of DeepSeek-V4.1-Flash behind the same OpenAI-compatible Chat Completions endpoint (https://api.unst33r.si/v1) and API key as the rest of the lineup.

Sitting between unst33r-flash and unst33r-pro on price, it fills a specific gap for security engineering: large-context code analysis, vulnerability reproduction, and fast multi-turn agent loops that need high output limits without mid-run refusals.

Specs and pricing

See the pricing page for the full lineup comparison.

Where DeepSeek-V4.1-Flash stands on public leaderboards

As with every reduced-refusal tier on UNST33R, public benchmarks evaluate the unmodified upstream base model rather than the abliterated weights (what abliteration changes). Even with that caveat, independent runs on Vals AI (October 2026) show why DeepSeek-V4.1-Flash is a strong base for security and code-heavy workloads:

Three details from the Vals AI runs matter in practice:

1. Vulnerability PoC & patching (CyberBench v1.1): CyberBench hands an agent a vulnerable open-source source tree and fuzz target binary, grading whether it can craft raw input bytes that trigger a sanitizer crash (PoC track) and patch the source without breaking holdout tests (Patch track). DeepSeek-V4.1-Flash scores 73.69%, finishing runs in roughly half the wall-clock time of GLM-5.3 (13m50s vs 30m18s).
2. Tool and skill execution (SkillsBench): On Vals AI's SkillsBench, DeepSeek-V4.1-Flash ranks #1 of 35 models (69.80%), making its base model particularly suited for harnesses that inject structured skill definitions or external tools.
3. 384k output ceiling: While most models cap output at 64k–128k tokens, unst33r-deepseek supports up to 384k output tokens in a single completion—useful when generating large diffs, full-file patches, or structured exploit test suites.

When to pick unst33r-deepseek vs Flash or Pro

For a full breakdown across all three available tiers, see which model tier for which security workload. The short rule of thumb:

• Use unst33r-deepseek for source-code auditing, crash PoC synthesis, large-diff patch generation, and high-speed agent loops where token efficiency and 384k output matter.
• Use unst33r-flash ($0.50 / $1.50) for everyday recon and interactive agent work at less than half the per-token rate.
• Use unst33r-pro ($2.50 / $4.50) for long-horizon interactive terminal sessions (where GLM-5.3 still leads on Terminal-Bench 4.0).

Quickstart

If you already use UNST33R in your script or agent harness (OpenClaw, Hermes, or the OpenAI SDK), switch the model parameter to unst33r-deepseek:

curl https://api.unst33r.si/v1/chat/completions \
  -H "Authorization: Bearer $UNST33R_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "unst33r-deepseek",
    "messages": [
      {
        "role": "user",
        "content": "Review this parser function for out-of-bounds read conditions and draft a minimal reproducer input."
      }
    ]
  }'

As with any abliterated model, we recommend running 5–10 prompts from your actual harness first to verify both completion quality and strict JSON/tool-call formatting under your workload. All usage is governed by the acceptable-use policy.

Get an API key to test unst33r-deepseek on your harness, or review the pricing page for token rates across the lineup.