FuturixAI open-sources CARVE, a Jev-like decision family on Qwen3.8-27B

What did FuturixAI publish?

The reader-facing announcement is FuturixAI’s DEV post dated 10 October 2026, which frames CARVE as an open contribution to structured decision-making. The implementing repository FuturixAI-and-Quantum-Works/carve-jev carries an Apache-2.0 LICENSE and a README that defines CARVE as Candidate-Aligned Readout via Vocabulary Evidence on Qwen3.8-27B. Repository history shows substantial branding and chart refresh commits on 9 October 2026; the git tree itself is older than the DEV post, so we date the public narrative to 10 October while noting earlier code activity.

CARVE’s “Jev-like” label is carefully scoped. The README and DEV post both say the phrase describes the typed decision interface and task category, and that CARVE is not affiliated with or endorsed by TypeSafe AI. That matters legally and editorially: compatibility with System One request shapes is an engineering claim, not a partnership announcement.

In the same week’s decision-model coverage, Microsoft’s hosted Decision-1 and open scorers such as THX-01 attack overlapping jobs—routing, classification, scoring—with different calibration stories. CARVE’s honesty about raw token preferences is part of what makes the release worth reading carefully.

How does CARVE make a decision?

Rather than sampling a free-form JSON object and parsing it, CARVE assigns candidate letters to user-defined options and scores only those vocabulary items. One output token decides among the restricted set. The stack then maps that choice into typed Noul, Choice, or Score answers through a System One-compatible API documented in the repo’s docs/ folder.

That approach sits between pure generative structured output and specialised decision heads. You keep a large Qwen3.8-27B backbone and vLLM-style serving, but you constrain the readout so downstream software receives stable keys instead of prose. The README emphasises prompt strategies, few-shot formats, caching, and adaptive demonstrations as levers—meaning operators should expect configuration search, not a single frozen prompt that wins everywhere.

Because probabilities are raw restricted candidate-token preferences, a 0.7 score is not a promise that the answer is correct 70% of the time. If your workflow needs calibrated deferral thresholds, you must fit calibration on your own validation set or choose a model that claims proper scoring-rule training. CARVE’s docs make that limitation explicit; ignore it and you will over-automate.

What do the Decision 1.0 numbers mean?

FuturixAI reports 82.22% accuracy on the public Decision 1.0 transfer-v9 benchmark—860 correct out of 1,046 decisions—from a local measurement. The same post cites Jev’s published 87.19% and Kev-9B’s 79.25% for orientation. Those competitor figures are externally published scores, not numbers FuturixAI re-ran on identical hardware in a locked harness. The DEV post states that distinction plainly.

Latency and throughput rows use an NVIDIA RTX PRO 6000 Blackwell Server Edition under warm serial inference and a concurrent load test: about 72.5 ms p50, 91.0 ms p95, and 20.23 requests per second across 5,120 requests with zero observed HTTP or response-validation failures in their run. Those are FuturixAI’s load-test notes. Your tokens-per-second will move with quantisation, batching, and prompt length.

If you already track open decision checkpoints such as Nace Drex v1.5 or Vega, keep protocols separate. Decision 1.0 transfer-v9, JevBench-style suites, and vendor ticket sets are not interchangeable leaderboards.

What is actually in the GitHub tree?

The repository FuturixAI-and-Quantum-Works/carve-jev includes quickstart material, API docs, a model card, reproducibility notes, and CI badges on the README we fetched. Apache-2.0 covers the project’s own code; Qwen3.8-27B weights remain under whatever license Alibaba attaches to that backbone—read both before commercial redistribution.

FuturixAI emphasises publishing protocols, frozen revisions, raw summaries, ablations, and negative results. That reproducibility stance is more valuable than another glossy accuracy claim. If you file issues or PRs, pin the revision you measured; README banners and chart commits on 9 October show the docs surface moves even when the core method does not.

Serving assumptions matter. CARVE’s path expects a generative model server that can expose logits or restricted decoding for candidate letters. Teams without GPU budget for 27B dense inference will not get the published latency profile from a laptop CPU. Evaluate whether a smaller specialised decision head meets the same product need before committing to CARVE’s backbone size.

When is CARVE the wrong tool?

Choose something else if you need calibrated probabilities out of the box, if you need native number/excerpt extraction with character offsets, or if you cannot host a 27B-class model. CARVE is also the wrong headline if you only need free-form JSON from a chat API—the point of the release is constrained candidate readout.

Choose CARVE when you want an open, inspectable System One stack, when you already standardise on Qwen3.8 serving, and when you are willing to run your own calibration and prompt ablations. The DEV post’s invitation to challenge findings is not filler; the benchmark margins versus published Jev scores are narrow enough that harness details will dominate arguments.