Darwin-27B-ZTC-v2: VIDRAFT's Open-Weights Answer to the Typed Decision Model Problem in Agent Graphs

TL;DR: TypeSafe AI's Jev raised $870M to prove that agent decision nodes don't need a chat LLM — they need structured, probabilistic outputs. VIDRAFT is positioning its open-weights model, Darwin-27B-ZTC-v2, as the Apache-2.0 alternative, claiming the top spot on the System One Mosaic Benchmark decision-engine leaderboard. If you're building agent graphs and tired of parsing routing logic out of generated prose, this architecture pattern is worth understanding.

What it is

The core thesis, validated by a $7.5B-valuation fundraise, is simple: most decisions inside an agent loop don't require natural language at all. TypeSafe AI's Jev model accepts a state plus a set of typed questions and returns calibrated probabilities — no sentences, no parsing, no ambiguity.

VIDRAFT's Darwin-27B-ZTC-v2 enters the same design space as an open-weights alternative. Key facts from the source:

• Model name: Darwin-27B-ZTC-v2
• License: Apache-2.0
• Claimed leaderboard position: #1 of 102 models on the System One Mosaic Benchmark decision-engine leaderboard
• Borda score: 89.58
• Task average: 66.46
• Approach name (partial): Zero-Token Calibration (ZTC) — the full method description was cut off in the source

The framing VIDRAFT is using is explicit: this is an open alternative to a proprietary typed-decision model, targeting the same agent-graph decision-node use case.

How it works

Understanding why this architecture matters requires stepping back from how most developers currently use LLMs in agents.

In a typical LLM-powered agent graph — say, a LangGraph state machine — routing nodes often work by prompting a chat model and then parsing its text response to figure out which edge to traverse. That's slow, expensive, and brittle. You're spending full generative inference budget on a question that has a finite, predefined answer set.

The typed-decision model pattern flips this:

• Routing nodes ask a Choice-type question whose options are the graph's transition labels. The model returns an explicit value, not prose to be interpreted.
• Policy nodes (e.g., safety checks) ask a Noul-style question — a binary statement — and receive a probability between 0 and 1. Application code then applies its own threshold logic:

  if probability >= 0.9 → execute
  if probability >= 0.6 → human review
  else → blocked

The model provides the judgment; the application owns the policy.

• Scoring nodes rate along an ordered scale when ordinal judgment is needed.

This keeps non-decision logic (null checks, input validation, length limits) in ordinary application code where it belongs — no model call needed for if (!input).

VIDRAFT's Zero-Token Calibration (ZTC) method (name confirmed by the source, full mechanism not disclosed in the coverage) is described as the approach behind Darwin-27B-ZTC-v2's performance on the decision-engine benchmark. The "ZTC" naming suggests the model may be optimized specifically for decision-type outputs without consuming output tokens in the traditional generative sense, but the source does not confirm implementation details — treat that as inference, not fact.

Benchmarks & results

From the source article, VIDRAFT's claimed public benchmark results for Darwin-27B-ZTC-v2 on the System One Mosaic Benchmark decision-engine leaderboard:

Metric	Value
Leaderboard rank	#1 of 102 models
Borda score	89.58
Task average	66.46
License	Apache-2.0

The System One Mosaic Benchmark is described as a decision-engine leaderboard, which means it is specifically measuring the model's suitability for agent decision nodes — not general instruction following or coding tasks. Developers evaluating this for general-purpose use should treat these numbers in that scoped context.

No other benchmark numbers, ablations, or comparative figures were included in the source coverage.

How to try it

The source article identifies Darwin-27B-ZTC-v2 as an Apache-2.0 open-weights model, which implies public availability — however, the article does not explicitly confirm a Hugging Face repository URL, a GitHub link, or an OpenAI-compatible API endpoint at time of writing.

Given the Apache-2.0 license claim and the "open-weights" framing, the model is likely accessible via Hugging Face once the release is confirmed publicly. Once a repository is confirmed, typical access would look like:

# Example only — verify the actual repo name on Hugging Face before running
huggingface-cli download vidraft/darwin-27b-ztc-v2

Do not use the above command until you have verified the correct repository name from VIDRAFT's official Hugging Face page or GitHub. The model name above is taken directly from the source but the exact repository path has not been confirmed in coverage.

Check VIDRAFT's official channels for confirmed access instructions.

FAQ

Q: Why would I use a typed-decision model instead of just prompting GPT-4o with a structured output schema?
A: Cost and latency, primarily. At $0.042 per million input tokens (Jev's published price), micro-decisions that fire hundreds of times per agent run become economically viable to automate. A full chat model call for every routing decision is overkill when the answer space is already enumerated. The same logic applies to open-weights alternatives — you can self-host Darwin-27B-ZTC-v2 and dedicate it specifically to decision nodes without burning your general-purpose LLM budget.

Q: Is Darwin-27B-ZTC-v2 a drop-in replacement for Jev?
A: Architecturally they solve the same problem, but the source does not confirm API compatibility. Jev exposes specific typed question primitives (Choice, Score, Noul) as a documented interface. Whether Darwin-27B-ZTC-v2 exposes the same interface or a different one is not confirmed in current coverage — check VIDRAFT's documentation before assuming interoperability.

---

Originally reported by 당무아(베트남) (2026-10-12) — source article.