AI Agent StoreFINAL-Bench · Open-weight decision checkpoint
Darwin 27B ZTC v2
This second Darwin release adds training for choosing actions and workflow steps. Test it separately from version 1, including whether its confidence scores match how often it is right.
First reviewed 2026-10-10 · Checked 2026-10-15. The repository was created October 8; a separate provider launch date is unverified.
We still need comparable independent tests. You can compare the confirmed features and limits below, but we cannot yet give this version an Agent Fit or Practical Value rating. Test it on records with known answers, including cases that should go to a person.
What you can use today
You download the model files and supply the hardware and software to run them. Capacity is the amount of information a request can contain; settings and available memory can limit what fits.
Exact checkpoint ID: FINAL-Bench/Darwin-27B-ZTC-v2. The supplied interface supports yes/no, choice and ordered-score questions without generating answer tokens. It uses a text-only readout head on the Darwin backbone; use the matching inference code and checkpoint.
The card declares Apache 2.0; bundled autojev code carries its MIT notice. The provider describes about 54 GB of BF16 weights, before runtime memory. v1’s card corrects mixed-run numbers. v2 discloses training on other splits of tasks that also appear in S1MB; we have not audited that leaderboard or imported its scores.
The configuration lists 262,144 internal positions, but a validated maximum for the decision interface is unverified. The v2 benchmark used an explicit 32,768 limit for one task; that is a test setting, not a universal endpoint allowance.
No hosted per-token price is verified for this exact checkpoint. Local GPU, memory and operating costs need your own measurement.
Provider release and specifications ↗
Check availability and plans ↗
Test it before switching
Give each model the same examples, tools and definition of success. Record correct outcomes, response time, retries and total cost. Include the time and hardware needed to run it.
Compare caching and long-context costs ↗