Jev directory
CommunityJev Models & Reimplementations
OpenThai-SystemOne Core AI
A Core AI conversion of OpenThai-SystemOne for Apple silicon: int8lin and fp16 .aimodel bundles with a 4,096-token context and 256-wide slot logits, gated against the author's fp32 oracle.
- Reading time
- 2 min
- Published by
- Community
- Author
- mlboydaisuke
- Last verified
- Oct 2026
Used inJev for Software & DevTools
Highlights
- Two bundles: int8lin is the ship build and fp16 the reference, both with a 4,096-token context and logits of shape [1, 1, 256].
- Option argmax matched the author's fp32 oracle on 50/50 question calls for both bundles, and the state reset left row 1 logits bit-identical.
- Max absolute probability difference: 0.020813 for int8lin and 0.005059 for fp16, with mean per-row differences at 0.000659 and 0.000225.
- The 18-request fixture covers Thai, English, mixed text, dict and list states, and rows with 40 and 255 options.
- Temperatures derived from exp(log_temperature): choice 1.058534, score 1.043141, noul 1.006767.
Watch out
Apache-2.0, converted from iapp/OpenThai-SystemOne revision f3709948 on Qwen3.5-0.8B-Base. Its tensor-derived temperatures differ from the author's rounded card values (choice 1.055, score 1.008, noul 1.047). The author's one-pass API can differ from independent-row answers, up to a 0.375453 probability gap, recorded as API behavior rather than a conversion failure.
Questions about this entry
Short answers drawn from the linked project or page.
int8lin is the ship bundle and fp16 is published beside it as the reference; both keep a 4,096-token context and emit logits with shape [1, 1, 256].
Option argmax equals the author's fp32 oracle on 50 of 50 question calls for both bundles, and on rows with oracle margin at least 0.02 it matches 49 of 49; probability differences are recorded rather than gated.
The card records a maximum absolute difference of 0.020813 and a mean of per-row mean differences of 0.000659 for int8lin, against 0.005059 and 0.000225 for fp16. The largest int8lin difference was a two-option noul row with oracle margin 0.061681.
The checkpoint reads exp(log_temperature) from the tensors at full precision: choice 1.058534, score 1.043141, and noul 1.006767, while the author's v0.3 card quotes rounded values choice 1.055, score 1.008, noul 1.047.
The author's per-question row layout with <|ts_state|>, <|ts_q|>, option tokens, and <|ts_answer|>; the bundle consumes the prefix through answer token id 248082 and reads the last call's 256 logits. For k options, mask slots k..254 and softmax over all slots.
The author's one-pass API places several questions in a single causal sequence, and answers to later questions can differ from independent rows by up to a 0.375453 probability gap. The independent-row oracle assembly equals the author's single-question API on 18/18 requests and 50/50 question calls.
The card's throughput figures are for a graph that is S=1, so it reports a prefill-rate proxy, not decision latency: 252.7 tok/s for coreai-pipelined and 197.1 tok/s for coreai-sequential, measured on an M4 Max GPU.
Reactions & coverage
Posts, threads, and videos about this entry from around the web.
X: Standard One SH (StandardThinking)
LinkedIn: What multimodal decision models actually change, and what they do not.
LinkedInJevDirectory.org
Hacker News: Decision Index 0.2.1 scores open reproductions of Jev on a frozen suite.
jackbravo4 pointsDecision Index 0.2.1 – 70 open reproductions of TypeSafe Jev's Decision ModelRead the thread on Hacker NewsHacker News: Pointer to a DiffusionGemma plus vLLM tutorial for a Jev-shaped classifier.
1 pointHow to build a Jev-style classifier with DiffusionGemma and vLLMRead the thread on Hacker NewsMore Jev resources like this
Fluid Inference's fixed-shape Core ML conversion of the classification path of Fastino's GLiNER2.5-Decide for Apple silicon: the DeBERTa-v3-large encoder and shared label classifier score up to four decision heads per call, in fp16, 8-bit or 6-bit packages.
Jev Models & Reimplementations#community#open-models#local
Community
Fastino's 1B English member of the GLiNER2.5 classification family: pass any label set at call time for intent, routing, sentiment or multi-label tags and get labels back in a single forward pass with no generated tokens. It averages 59.6% exact-match on Fast Decisions.
Jev Models & Reimplementations#community#open-models#local
Community
Fastino's 340M English classification model and the headline member of the GLiNER2.5 family: pass any label set at call time and get intent, routing, sentiment or multi-label tags back in one forward pass, at 60.2% exact-match average on the Fast Decisions suite.
Jev Models & Reimplementations#community#open-models#local
Community
Mohit Garg's Apache-2.0 vision-language decision model on Qwen/Qwen3.5-4B, the recommended default of the imajev family: it reads photos against records and text, returning calibrated probabilities and a trained can't tell.
Jev Models & Reimplementations#community#open-models#vision-language
Community
TokenRhythm's 4B prefill-only decision model built on NeoHorse-1-4B: given a state and application-defined questions it returns Choice, Noul, and Score probabilities without generating text, reporting a 77.70 six-group aggregate under Apache-2.0.
Jev Models & Reimplementations#community#open-models#local
Community
An unofficial cross-encoder fine-tune of ModernBERT-ja-310m for Jev Local: it reads a question-plus-state and one candidate per sequence and hits 92.62% on the JGLUE JNLI test, with CC BY-SA 4.0 weights.
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.