Jev directory
CommunityJev Models & Reimplementations
THX-01
HAL-X AI's 322M non-autoregressive decision model returns calibrated answers in one forward pass of about 10 ms, with native number, excerpt and citation outputs beyond Jev's choice, noul and score types. Apache-2.0, post-trained in 18 languages.
- Reading time
- 2 min
- Published by
- Community
- Author
- HAL-X AI
- Last verified
- Oct 2026
Highlights
- 322M parameters: a 307M mmBERT-base encoder (22 layers, 256k vocabulary) plus a 15M decision head; 1,024 tokens per question.
- Native number, excerpt and citation answers beyond Jev's choice, noul and score, which need repeated choice calls in the service layer.
- Trained with eight large-scale RLCD stages over about 2.1 million decisions, rewarding a strictly proper scoring rule.
- Support-ticket benchmark: 98.4 average accuracy over four sets, ECE 0.003, about 10 ms per request; Jev 1.13 is listed at 97.4, ECE 0.007, 331 ms.
- Number lookup 93.4 and excerpt extraction F1 84.1 versus 69.6 and 20.1 for the previous checkpoint; Jev 1.13 reaches 96.6 on number lookup.
Quickstart
python
import thx01
agent = thx01.load("doofz/THX-01") # GPU if available
doc = ("Northwind Corp. reported Q3 2026 results. Revenue for the quarter was $48.3 million, up 14%. "
'"We will open two offices in Baku," said CEO Laura Chen.')
result = agent.decide(doc, {
"kind": {"type": "choice", "question": "What is this document?",
"criteria": {"earnings": "earnings report", "complaint": "customer complaint", "invoice": "invoice"}},
"revenue": {"type": "number", "question": "What was the revenue, in US dollars?", "cite": True},
"quote": {"type": "excerpt", "question": "What did the CEO say?"},
})Watch out
Model card reviewed 2026-10-15; the fetched text was truncated inside the calibration section. The benchmark tables are the project's own, Claude Sonnet 5.5 was evaluated on a stratified 200-ticket subset, and LLM baselines return no probabilities so their ECE is not shown.
Questions about this entry
Short answers drawn from the linked project or page.
A non-autoregressive, multilingual decision model from HAL-X AI with 322M parameters: a 307M encoder plus a 15M decision head. Given a state and typed questions, it returns a calibrated answer to every question in a single forward pass of about 10 ms on one GPU.
It follows the same typed-decision idea: choice, noul and score match TypeSafe's question types, and THX-01 adds native number, excerpt and citation answers that Jev handles through repeated choice calls in the service layer.
choice returns one of N options with per-option probabilities, noul returns P(yes), score returns an ordinal level and distribution, number returns a value stated in the document or null, and excerpt returns a verbatim span with character offsets.
Across four test sets of 2,843 tickets in Azerbaijani, Russian, English and Turkish it averages 98.4 accuracy with ECE 0.003 at about 10 ms per request; Jev 1.13 is listed at 97.4 average, ECE 0.007 and 331 ms.
Reinforcement Learning for Calibrated Decisions: the reward is a strictly proper scoring rule, so reporting honest probabilities is the only way to maximise it. Post-training ran eight stages over about 2.1 million decisions.
pip install thx01 for the library, or pip install "thx01[server]" for the REST server. Weights download from the model repository on first use, and the licence is Apache 2.0.
Input is capped at 1,024 tokens per question and longer states are truncated. number performs no unit or currency conversion, returning null when the document states a different unit, and the benchmark tables are the project's own.
Reactions & coverage
Posts, threads, and videos about this entry from around the web.
X: Both Jev 1.13 and 1.14 score 29 of 30 on 30 Turkish buyer and seller messages
X: A 26-sheet plan set in 2.9 seconds
X: Which exam questions come back
X: 900 images in 40 seconds
More Jev resources like this
Gestalt Labs' Jeff is a LoRA adapter on Qwen3-4B-Instruct-2507 that answers Jev-style choice, noul and score questions with labels and probabilities instead of prose, runs locally, and ships a train-your-own guide; on 9,730 fact-checking examples it trails Jev 1.13.0 on accuracy (0.8183 vs 0.8283).
Jev Models & Reimplementations#community#catalog-import#madewithjev
Community
A CPU port of SemIf with a web UI: JEV-CPU loads Qwen3-0.6B in float32 and reads a decision from its option logits in about a second, across eight domains from support to incident severity, with no GPU.
Jev Models & Reimplementations#community#catalog-import#madewithjev
Community
Zefan Cai's LoRA adapter and scalar decision head for Qwen3.5-9B score caller-supplied candidates for choice, yes/no and score answers without generating text, served over a local /v1/systemone endpoint.
Jev Models & Reimplementations#community#catalog-import#madewithjev
Community
An open-source, Jev-compatible System One model: Dex answers choice, score and yes/no questions over text in a single forward pass from an Ettin-1B encoder, and speaks the same /v1/systemone shape as Jev.
Jev Models & Reimplementations#community#catalog-import#madewithjev
Community
canadaornaw/dexPython
canadaornaw/dexPythonThe v2.2 PEFT LoRA adapter for Standard Thinking's Standard One 3B decision model, shipped with a merge recipe for Ministral-3-3B-Instruct; serving requires the merged checkpoint and the shared server code in the StandardOne-8B repository.
Jev Models & Reimplementations#community#open-models#local
Community
Standard Thinking's 8B LoRA adapter for the Standard One decision model: merge it with the supplied recipe and serve it, and it scores choice, noul and score labels over a scenario with calibrated probabilities via POST /v1/systemone, never generating text.
Jev Models & Reimplementations#community#open-models#local
Community
From the Jev guides
Original write-ups on what this entry is for.
From the Jev AI model community
Posts from builders shipping with Jev right now.