Updated
Topic
Open source
Open-weight stand-ins and local ports, with the numbers their authors published. 84 stories.
OR-decision-4B prints JevBench 193 of 231, and Choice 79.7 against Jev 79.6 on alphaXiv's chart
alphaXiv posted OR-decision-4B on October 30, 2026, at 16:06 UTC. The Hugging Face card prints JevBench 193 of 231, 83.55%. The launch chart prints Choice 79.7 against Jev 79.6, with sarcasm at 53.6 against Jev 70.7. Apache-2.0 LoRA on Qwen3.5-4B. This desk did not run the weights.
BranchOne returns Choice probabilities from Qwen3.5-0.8B-Base, median 0.305 s on ten short text questions
Necati Demir posted BranchOne on October 29, 2026. The open checkpoint returns one probability per supplied candidate from Qwen3.5-0.8B-Base plus LoRA. His A100 Colab run prints median 0.305 s on ten short text questions. Apache-2.0. This desk did not run the weights.
MiniCPM5-2B-Jev prints 78.79% on JevBench, first among open models at or under 2B
OpenBMB posted MiniCPM5-2B-Jev on October 28, 2026, at 13:01 UTC. The project prints 78.79% on JevBench, first among open models at or under 2B, and 86.41% accuracy with ECE 0.0307 on decision-v7. Apache-2.0 on MiniCPM5-2B. This desk did not run the weights.
Opened October 26, Decision Index 0.4.1 Full score for d3-ultra is 70.18, second of 229
Xunzhuo Liu posted d3-ultra on October 27, 2026, at 03:01 UTC. Opened October 26, Decision Index 0.4.1 Full score is 70.18, second of 229, +3.03 over d3 at 67.15. The card's private SR-Text mean is 87.2. Apache-2.0. This desk did not run the weights.
hunch runs EmbeddingGemma2 as a CPU System One endpoint at 20 ms on text and 100 ms on images
Matt Mastracci's hunch turns EmbeddingGemma2 into a CPU System One endpoint at 20 ms on text and 100 ms on images. Cloudflare Clef teaches a local student in real time. Open-source launch was planned for later the same week. This desk has no public repo or hosted call yet.
Base Labs posts mammoth-1 at Decision Index 0.4 Full 71.24; live 0.4.1 prints 68.4
Base Labs posted mammoth-1 on October 21, 2026, at 17:14 UTC, and the card prints Decision Index 0.4 Full 71.24, first of 208. Opened October 23, the live space is edition 0.4.1, 229 models, mammoth-1 Full 68.4 rank 3. Date headers 02:23:21 GMT to 02:23:23 GMT, GET /v1/models with no key returned HTTP 401.
Sthānika's Sieve-31B card prints Decision Index 0.3.2 Full score 63.8
Cuth posted Sieve 31B on October 20, 2026. The Hugging Face card prints Decision Index 0.3.2 Full score 63.8, board data of October 14, and a public index of 62.52 on the authors' own 0.3 run. Apache-2.0 LoRA on Gemma 4 31B. This desk did not run the weights.
Opened October 26, Decision Index 0.4.1 Full score for d3-turbo is 67.31, and vision full is 72.46
Xunzhuo Liu posted d3-turbo on October 20, 2026, at 00:42 UTC. Opened October 26, Decision Index 0.4.1 Full score is 67.31 and vision full is 72.46, second of 56. The card's own 0.3.2 estimate is 70.2. Apache-2.0. This desk did not run the weights.
ToddlerJev prints 68.70% on MMAU-mini, with weights under the Qwen Research License
ToddlerJev, posted October 18, 2026, prints 68.70% on MMAU-mini, 61.30% on MMSU, and 56.90% on MMAR. Weights are TeenageWorks/Toddler-Jev under the Qwen Research License. This desk did not run them.
Opened October 19, JevBench ranks basal-omni-1.0-ultra first of 184 open systems on Capability at 82.5
Opened October 19, 2026, JevBench ranks basal-omni-1.0-ultra first of 184 open systems on Capability at 82.5. Composite among open weights is 71.8, rank 3. The card prints Werdykt v1 at 0.879. This desk did not run the weights.
Span-02's card prints Decision Index 0.3 text index 63.62, and the blog prints thinking off at 61.2 against Jev 58.0
Hendrix Liu posted October 16, 2026, 23:16 UTC, that Span-02 had more usage in under 24 hours than Span-01 did in 20 days. The card still prints Decision Index 0.3 text index 63.62. The blog prints thinking off at 61.2 against Jev 1.13.0 at 58.0. This desk did not run the weights.
Jeva-4B prints 0 flips on 600 Choice permutations, and BANKING77 at K=16 is 91.62 against Jev 82.50
Hugging Face DailyPapers posted Jeva on October 14, 2026, at 20:10 UTC. The paper prints 0 flips on 600 Choice items and BANKING77 at 16 candidates 91.62 against Jev-1.13.0 at 82.50. Apache-2.0 on Qwen3.5-4B. This desk did not run the weights.
Mizan 4B averages 78.7 on 3,000 Persian records
xmanii/Mizan-4B prints a mean accuracy of 78.7 on 3,000 real Persian records, against the card's Jev cell of 83.7. Clef 27B on that table is 83.4 and Laya is 62.4. The Q4 GGUF mean is 78.3 at 2.5 GB. The card says the weights are Apache-2.0, from Qwen3.5-4B, and text only. We did not run the weights.
THX-01 averages 98.4 on its 2,843 support tickets
HAL-X's THX-01 card averages 98.4 on 2,843 support tickets, against Jev 1.13 at 97.4 on the same four sets. Number lookup on a different set is 93.4, and Jev on that line is 96.6. The card says 322M parameters, Apache 2.0, on mmBERT-base. The October 11 post calls the API free. The docs page we opened does not print a price. We did not call the API.
PurpleMIST-Flash-1.0 scores 0.668 on 15,050 translated African-language decisions
PurpleMIST-Flash-1.0 scores 0.668 accuracy on 15,050 translated African-language decisions, against DeepSeek-V4-Pro at 0.669. KL on that set is 0.214 against 1.125. The card says the checkpoint is Apache-2.0, about 7.9B parameters, from Qwen3.5-9B-Base, and text only. A self-run of Decision Index edition 0.3 prints a public index of 46.7. We did not run the weights.
TypeLLM scores Qwen3.8-27B at 228 of 231 public JevBench tasks when thinking is on
On October 16, 2026, at 05:44 UTC, TypeLLM posted that Jev 1.14 cannot return an array of vehicle boxes, including Tesla's flying Roadster. The cookbook at 14.00 s prints six boxes. The September 23 file still scores Qwen3.8-27B at 228 of 231 public JevBench tasks with thinking. This desk did not call the API.
Vega, the vegaml package, scores 75.4% on 800 phishing emails against Jev at 61.9%
Nandha Kishor M's Vega, the vegaml package, scores 75.4% on 800 phishing emails against Jev 1.13.0 at 61.9%. The default checkpoint is 0.8B on frozen Qwen3.5-0.8B, Apache 2.0, with a 4B sibling. Decision 2.0's Vega is a different checkpoint. We did not run the weights.
Clef-omni scores a 21-second video in about 1.5 seconds, at $0.15 per million input tokens
Cloudflare's October 9, 2026 blog prices Clef-omni at $0.15 per million input tokens. The same post cuts hosted Clef-flash from $0.09 to $0.038 and from a 64k window to 24k. A 21-second video with sound scores in about 1.5 seconds. On that post's ten-benchmark table, Clef-omni's BANKING77 macro-F1 is 94.8. The weights are Apache 2.0, from Qwen3-Omni-30B-A3B-Instruct. We did not call the API.
OpenBles Firewall 1 catches 119 of 120 agent attacks and flags no benign look-alike
OpenBles Firewall 1 caught 119 of 120 agent attacks and flagged none of 120 benign look-alikes on its held-out test. The README prints classic-set recall at 55 of 60, with 2 of 56 benign prompts flagged. The adapter is MIT, 31.8 MB, on fastino/GLiNER2.5-Decide. We did not run the weights.
Open decision models start from an open language model
Open decision checkpoints start from a public language model because a closed API does not release the weights a decision head has to read. Jeff v1.3 prints a mean of 30.4% for untrained Qwen3.5-0.8B and 92.8% with an adapter, across 13 of its own tests.
Step-Jev scores 19.6% when the reward grades every search step
On October 9, 2026 hosted jev-1.13.0 scored the repository's no-step test state incorrect at probability 1.0. The five-step fixture in the same test was correct at 0.05. On a Moon-landing question written here, a no-search answer of 1965 was incorrect at 1.0, and a repeated NASA search stayed positive at 0.53. real_jev_gate.py's rollout files are not in the repository.
H2O-Lightning-4B v1.1 prints JevBench composite 72.5 against Jev at 71.5
H2O's October 7, 2026 blog prints H2O-Lightning-4B v1.1 at JevBench composite 72.5, against Jev 1.13.0 at 71.5. That reading is board revision v1.7.17 and 104 open-weight systems. The blog says the weights are Apache 2.0, from Qwen3.5-4B. We did not open the card, and we did not run the weights.
CLERC top-1 moves from 5% to 18% when jev-1.12 reranks the shortlist
TypeSafe's CLERC cookbook moves top-1 from 5% to 18% when jev-1.12 reranks a BM25 shortlist of 30 on 40 queries. Hindsight's listwise rerank later prints LoCoMo recall@1 0.950 against MiniLM at 0.800. aifabrice's best NFCorpus row is nDCG@10 0.451. We did not rerun them.
Open d1's d1-3B card prints 48.57 on a self-run of Decision Index 0.2.1
On October 7, 2026 Liquid AI posted Open d1, and the d1-3B card prints 48.57 on a self-run of Decision Index 0.2.1. Winnow-12B on that table is 50.02. d1-omni-600M is 15.95. The LICENSE file is LFM Open License v1.0, with commercial use under a $10,000,000 revenue line. We did not run the weights.
PolicyLM-1.7B scores each policy category from 0 to 1
PolicyLM-1.7B prints custom-policy accuracy 0.842 at a 0.5 cutoff, and it returns a score from 0 to 1 for each category, up to 16 in one pass. The weights are Apache-2.0. The table has no Jev column.
Nace posts Drex 1.1, and the card titles the weights Drex DLM
Nace posted Drex 1.1 on October 6, 2026, and the card scores Decision Index 0.2 at 52.31 against the kit's Jev cell of 51.67. Weights are CC BY-NC 4.0. Nace's code is MIT. The call is POST /v1/systemone. The hosted Drex 1.5 row on Decision Index 0.2.1 is 58.28.
Blink v0.3 prints Decision Index 0.2.1 at 57.48
Pixilab's Blink v0.3 card prints a self-run of Decision Index 0.2.1 at 57.48, against the board's Jev cell of 57.91. The card's license field is Apache-2.0. The base is Gemma 4 26B-A4B. We did not run the weights.
Quyet-1.0-Large's homepage prints JevBench v1.6.0 at 71.39
Quyet-1.0-Large's homepage prints JevBench v1.6.0 official score 71.39, rank 1 of 92, and Capability 81.7. The card says Apache-2.0, 31.3B, from Gemma 4 31B. A later reading on this site still lists Capability 81.7 and official composite rank 3. We did not run the weights.
AutoTrust scores GEV-26B-Decide at 62.48 on its own Decision Index run
AutoTrust posted GEV-26B-Decide on October 3, 2026, a Gemma 4 26B model with about 4B active. The card's own Decision Index 0.2.1 scoring prints an adaptive 62.48 against the September 28 board's Jev cell of 57.91. Adaptive thinking on Knowledge misses the kit's latency cap. A 1,754-question holdout moves from 73.3% to 83.4% when thinking is allowed. The project is not affiliated with TypeSafe.
NIRNAY 450M prints 0.8792 on 3,080 Banking77 rows and says the Jev cell was not measured here
Gautam Kishore posted NIRNAY 450m on October 3, 2026, an Apache-2.0 Laya fine-tune from Eulogik. The phase_b JSON is 0.8792 on 3,080 Banking77 test rows, Brier 0.208, fitted ECE 0.045. The post's headline 0.8656 is the phase_a file, set against a Jev figure of 0.64. The card also prints Jev at 0.803, and the post says that column was not measured in the repo. The eval JSON has no Jev field.
WaterSheep fine-tunes ModernBERT and serves /v1/systemone with a multi-label question
Samrat Dutta posted WaterSheep on October 3, 2026. The Apache-2.0 repo fine-tunes ModernBERT-base and serves POST /v1/systemone on port 8766, with noul, choice, score, and multi. The README prints 77.8% on an in-distribution split, ECE 0.026, and 61.2% on held-out datasets, ECE 0.043. Those two rows do not print a question count, and the page has no Jev column. We did not run it.
Cloudflare posts Clef and Clef-flash, Apache 2.0 decision models on Workers AI
On October 9, 2026 Cloudflare's blog prices Clef-flash at $0.038 per million input tokens, down from the $0.09 on the October 1 card. Clef stays at $0.24. Clef-omni, posted the same day, is a separate model at $0.15. Cloudflare posted Clef and Clef-flash on October 1, 2026, two Apache 2.0 decision models hosted on Workers AI and released on Hugging Face. The October 1 blog's median latency row is 209.3 ms and 38.8 ms against a printed Jev cell of 524.1 ms. On October 3 a Fastino chart prints Clef at 61.2 and Clef-flash at 57.1, and captions those Cloudflare scores as self-reported. llama.cpp v0.6.0, published October 5, serves Clef on /v1/systemone, including images. The October 6 note leaves Clef's speed cell blank.
Strands Decider 2B scores 167 of 231 public JevBench tasks
Strands Labs posted Decider 2B on October 1, 2026, scoring 167 of 231 public JevBench tasks at a median 115 ms on an RTX 3090.
The Jeeves README prints 0.865 on the public JevBench hard tier
PostHog/jeeves, MIT, is a Qwen3.5-9B LoRA plus a pointer head. The README's thinking run, greedy, 2,560-token cap, prints 0.865 on 111 public JevBench hard items, against a Kev-published Jev cell of 0.730. On the transfer overall, Jev is 0.800 and Jeeves is 0.746. The sealed judge tier is not in the JevBench rows. We did not run it.
Maincode's Matilda card prints 59.26 on a self-run of Decision Index 0.2.1
Opened October 8, 2026, the Matilda card prints Decision Index 0.3 at 60.16, and the model API lists private false. Maincode posted Matilda on September 30, a 26.1B model under an Apache 2.0 line, and a self-run of the Decision Index 0.2.1 kit at 59.26 against the board's published Jev cell of 57.91. The card's median is 56.6 ms per request. The post's text says 56.8 ms. The run section still says the weights are private. We did not run the model.
Ollaya serves local decision models on port 11435, and prints winnow:e4b at 0.722
ollaya-dev/ollaya, Apache-2.0, calls itself Ollama for decision models and says it is independent of Ollama and of TypeSafe. The README prints winnow:e4b at 0.722 against TypeSafe Jev at 0.738, and 89 ms for five questions on an RTX 4090. The item count behind that pair is not defined on the page we read. We did not install it.
JevScout ranks oversized MCP text and prints 38% less agent time on its own check
Tim Lin posted JevScout on September 29. The MIT hook leaves Claude Code's MCP result alone under 100 KB, and otherwise asks Jev which segments to keep verbatim. The README's check is 38% less agent time and 26% less agent cost, with TypeSafe's bill separate. One printed row is 5 of 5 against 1 of 5. The evaluation note does not list all seven tasks. On September 30, 2026 this desk scored seven stand-in tasks, three times each, on jev-1.13.0. The critical line stayed in all 21 filtered runs. Claude Code was not installed, so the 38% and 26% were not measured.
A Pi extension asks Jev which skills to load, and the 20-skill demo drops about 1,000 tokens
Arnav Gupta posted a Pi skill picker on September 29. The MIT repo, pi-skill-picker-jev 0.1.0, asks a decision model which skill names belong in the prompt. The README's 20-skill storefront demo cuts the initial system prompt by about 1,000 tokens. The post's 250-skill work repo, about 10,000 tokens and under a second, is not in the README. We did not boot Pi. On September 30, 2026 we scored the 20 published skill names, and a 250-name catalog of those names plus 230 fillers written here. Necessary-skill recall was 16 of 19 on both catalogs. Five of eight tasks had every required skill selected. Median latency was 951 ms and 2,322 ms. The scoring note, opened September 30, uses a five-level Score and a default threshold of 0.625.
DiffusionGemma as Jev reads one denoising step, and a hand chart sums to 198 of 201
Google's Gemma account pointed at vLLM pull 57250 on September 29: seed a canvas, read yes/no, choice, and score probabilities in one denoising step. Matt Mastracci's September 17 chart adds to 198 of 201 against Jev at 191 of 201. A public 231-task replay is 196 against a published 200. On a 100-point routing file, djev-spark's joint score is 32.2% against Jev at 61.4%.
imajev-4b: Hugging Face weights and JevBench 70.4
Opened October 8, 2026, the JevBench v1.6.1 open-weights list puts imajev-4b at capability rank 31 and 61.9. Image JevBench v0.3.0, opened the same day, lists capability 68.0 at rank 1 and official composite 63.7 at rank 2 of 52. On October 7, 2026 the open-weights list put imajev-4b at capability rank 17 and 61.9. imajev-4b is Mohit Garg's Apache-2.0 LoRA on Qwen3.5-4B, with Hugging Face weights, JevBench v1.5.6 rank 10 at 70.4, and Image JevBench rank 2 at 76.4 on October 4. The September 27 changelog scores the same checkpoint 67.37 at rank 1, and Jev 1.13.0 stays 63.29 on that board. A September 26 adapter moved ImajevBench from 82.4% to 83.9%. We did not rerun the boards.
A 400M encoder named Kev scores 79.96% against Jev on 17,476 of its own rows
Sandipan Kundu posted skundu42/kev, a 400M encoder with Choice, Noul, and Score. On his held-out mixture Jev 1.13 scores 73.54% and this Kev scores 79.96%, ECE 0.015 against 0.098, median 132.7 ms against 388.4 ms. On a phishing set he did not retrain for, the same checkpoint flags 1,995 of 2,000 emails and accuracy is 50.25% against a published Jev figure of 62.6%. The root listing has no LICENSE file. This is not Jared Palmer's Kev. We did not rerun it.
Plumb-4B reaches 89 of 111 public hard items, and the official board scored it 65.84
Christopher House published crh225/plumb, Apache-2.0, starting from JevK5 v0.2. His runner, 231 public items on one RTX 4080 Super, puts v5 at 0.802 on the hard tier, 89 of 111, against the starting checkpoint at 0.739. The JevBench changelog for v1.4.2.1 scores the submitted model 65.84. On the v1.5 roster addendum it would place fourth at 71.6, outside the frozen ranked order. The v1.5.4 page opened September 30 numbers that same 71.6 as official rank 5, still tagged roster addendum A1. We did not rerun either table.
jev-code's three-task ladder still fails after the call-nesting fixes
Roberto Montalti posted rhighs/jev-code on September 26. Jev picks grammar productions. The program builds the AST and runs a compiler. The guide's ladder, one run per task, dated September 18, still fails guessing-game, file-io-script, and multi-file-package on commit 517d0f0. A hello-world smoke test does pass, in 3 turns and 16 requests. package.json is 0.2.0 and has no license field. The root listing has no LICENSE file. We did not run it.
jevos-1b prints 0.815 on 2,000 yes/no questions, against Jev at 0.927
Federico Elia released feder-cr/jev, a local server called jevos-1b, built with Loris Salsi. The README's image alt is the printed bench: on 2,000 yes/no questions from unseen policies, accuracy is 0.815 for jevos, 0.927 for Jev, and 0.489 for Laya. Short-prompt latency is 54 ms, 344 ms, and 104 ms. The server accepts Noul and returns 422 for Choice and Score. pyproject says MIT. The root listing has no LICENSE file. We did not rerun it.
Jevrail scores a shell command before it runs, and the printed corpus leaves the model off
Saman Pandey posted jevrail, an MIT hook that asks Jev whether a shell command is likely to destroy something, then blocks, asks, or allows it. The published eval is an 80-line corpus with the model switched off: 22 of 22 catastrophic commands are caught, and 15 of 26 safe commands still raise a false ask.
On Ámbar Reflex the 4B logit player went further than the more accurate 9B
Abel Coronado's local server plays an original run-and-gun with six typed questions and no generated text. Qwen3.5-4B at 8-bit on an RTX 3060 advances 24.9% of the level in 408 ms a decision. The same size model writing JSON advances 10.3% in 1,041 ms. Offline, Qwen3.5-9B is right on 82.7% of 554 labelled decisions and reaches 17.5% when the clock is running.
jev_local's Sarashina row beats LFM on two JGLUE tests, and the README says that is not a ranking
あるごす added Sarashina2.2 Vision 3B to a local /v1/systemone server. On JGLUE test the Q4 checkpoint is 34% on JNLI and 86% on JCommonsenseQA, against LFM at 17% and 69%. ModernBERT-Ja, trained on the JGLUE train split, is 93% and 92%. The post times the 3B model at about 100 ms when it is fast and about 500 ms as the usual case, on a Radeon AI Pro R9700.
Jevre's test split is 0.9323, and held-out sources fall to 0.8366
Di Zhang released a 0.85B checkpoint that speaks the TypeSafe Python SDK from localhost. On 12,000 evaluation rows the test split is accuracy 0.9323 and ECE 0.0079. Held-out sources are 0.8366 and ECE 0.0944. Replacing the state drops both near 0.43. The served path is 0.9267 at 77 ms median. The post calls the training RLCD. The README names Plackett-Luce ranking and a Brier term.
postgres-search can sink a bad hit, and the README has not measured that step
Alex posted an Algolia-style search in Postgres with an optional Jev pass. On 50 handwritten queries the SQL ranking puts a correct product first 82% of the time. The Jev step asks one Noul per candidate in the top 10 and sinks anything under 0.30. The docs say that pass was not measured, because no API key was available.
Rizzo Flow is 49 ms on an RTX 5060 Ti, and six missing-evidence rows stay confident
Simone Rizzo's Apache-2.0 server reads Spark-X2.5 logits through llama.cpp and speaks POST /v1/systemone. On Windows with an RTX 5060 Ti, the 4B Q8_0 checkpoint is 49 ms at the median and uses 5.6 GiB. On SemIf's authored144 fixture the balanced accuracy is 0.812 against SemIf's published 0.819. Where the right answer is insufficient evidence, 6 of 36 rows are still confident and wrong.
JevGate turns one function at a time into a finding at 0.80
Nodehaven Online posted jevgate, a Rust review gate on crates.io as v0.5.1. Tree-sitter finds a function, a copy, or a test locally. Jev answers a short question. Code maps the probability to review, consider, note, clear, or uncertain at 0.80. The README says those probabilities are judgments, not a measured accuracy.
Jevonian asks Jev which model should take the next coding turn
Chen posted xinyao27/jevonian, an AGPL-3.0 local proxy on port 8787. The virtual model jevonian/auto makes one Jev call for the route and the thinking depth. A named phase, a real model id, or routing turned off skips that call. The README prints no scored routing table.
A 100,000-line needle takes 18.0 minutes with Jev and 24.7 without
El Asombroso Nosequien posted elalber2000/jev-rlm, a fork of a minimal recursive-language-model loop. On one 100,000-line file both runs returned 2867825. results.json records 1,081 seconds with 82 Jev calls and 1,480 seconds with none. The post's dollar figures are an estimate. The file has no charge field.
Together prices Tev1 input at $0.04, and Ollama serves Choice, Score, and Noul
Opened October 7, 2026, Together's Tev1 page lists input at $0.04 per million tokens and free output. The September 23 post listed $0.042. The chat call returns one letter. Ollama serves tev1 on /v1/systemone.
AgentOS posts a 121-turn routing table for an opt-in Jev strategy
AgentOS added Jev as a third Pilot Router strategy, off by default. On the project's 121-turn labeled set the chart shows Jev at accuracy 0.901 against the local MiniLM classifier at 0.355. The same chart prices 121 cases at about $0.005 and a call at about 0.9 seconds. A follow-up pull request the same day says it sharpens the R2 and R3 criteria. Opened September 30, that pull's file list has no results JSON. Its changelog prints accuracy from 0.760 to 0.901.
A Cloud Run recipe serves DiffusionGemma-Jev at 35 to 60 ms
Google's Gemma account posted a one-command Cloud Run deploy for DiffusionGemma-Jev. The README puts single-request latency at 35 to 60 ms, batch throughput at 100 to 123 requests a second, cold start at about 47.5 seconds, and the GPU at $3.19 an hour while it is up. The repository lists no license. It does not print an accuracy table.
hotdog-27B leads a 1,631-decision holdout and trails on 74 public items
Michael posted hotdog-27B, a rank-16 LoRA on Qwen3.8-27B that reads two logits and returns yes or no. On an independent holdout it is 1,603/1,631 against Jev 1.13.0 at 1,587/1,631. On 74 native-binary public JevBench items it is 60/74 against 62/74. The second interval includes zero.
jev-gateway asks Jev which tool a coding agent should call
Vini Lana posted jev-gateway, an MIT local proxy for Codex, Claude Code, OpenCode, and Gemini CLI. Jev picks the tool. Closed arguments can skip the language model. On the project's chess-engine bench, GPT-6 Astra and GPT-5.6 Sol cut bug-fix output tokens by 57%. The feature task made Opus 5 and Sonnet 5 use more.
jev-rubiks asks Jev when to speak, after a test where picking moves did nothing
maxlibin/jev-rubiks is an MIT browser cube. Kociemba solves it in code, in about 20 ms. Jev only decides whether the coach should talk. The README says a removed experiment let Jev pick among 18 turns, forty times, and the scramble distance stayed at 22, the same as a random walk that went from 22 to 21.
jym asks Jev which real subcommand a mistyped one was aiming at
syumai posted jevyoumean, an MIT Go wrapper. When a subcommand is not in the CLI's help, it sends one Jev Choice over the documented names and descriptions, then offers the top matches. Valid commands never call the API. The README ships an eval harness and does not print a scored corpus.
jsort ranks a file by asking Jev which of two texts is more of something
Khaled Eltokhy published jsort, an MIT command that fits a Bradley-Terry scale from pairwise Jev judgments. On 300 CommonLit excerpts it tracked teacher easiness at Pearson r 0.824 for about five cents. Joe Weisenthal then said he had rescored 4,005 FOMC speeches and statements with it. The repo's own Fed check is 95 opening statements.
laya-mlx runs the Laya checkpoints on Apple Silicon
mizorewww/laya-mlx is an Apache 2.0 MLX port of Convai's Laya weights. On an M3 Max the README's one-question median is 13.42 ms for the 421M checkpoint and 7.39 ms for the 322M multilingual checkpoint. Selected answers matched upstream on 63 of 63 validation questions. The port does not publish a new accuracy table against Jev.
llamacpp-jev answers /v1/systemone from an unmodified llama-server, including images
Chirag posted llamacpp-jev, an MIT wrapper that puts TypeSafe's decision API in front of stock llama-server. On an M4 Pro, four questions about a fresh 448 by 448 image took 526 ms median, 32 of 32 on synthetic shapes. The probabilities are a raw label softmax. Qwen3.5-2B picks option 16 on every 64-way question the notes tried. A reply the same hour as the cited 800 ms post links kishida/llama.cpp branch jev, which builds the endpoint into llama-server. On October 2, upstream llama.cpp merged the same path for five decision-model GGUFs. Release v0.6.0, published October 5, also serves Nimble and Clef. The Nimble pull names Bespoke-Nimble-9B-v3, whose card prints 56.88 on Decision Index 0.2.1. Clef can take an image. The October 2 medians are unchanged.
rev reads a frozen Qwen and posts 0.784 on public hard items
Steven Yu posted 54yyyu/rev, an MIT harness that reads option logits from a frozen open model. On the public JevBench tiers, Qwen3.8-27B through sglang scores hard 0.784. The same README prints Jev 1.13.0 at 0.730, and 0.721 on a September 22 measurement. A laptop Qwen3.5-2B scores hard 0.550.
spark-s1-4b-v6 swaps in Qwen3.5 and prints a public-tier 83.1
Abhishek Rai posted spark-s1-4b-v6, an Apache-2.0 LoRA on Qwen3.5-4B. The README's public JevBench proxy is Intelligence 83.1, hard 0.595 of 111. On the v1.5 page opened September 29 the same checkpoint is official rank 7 at 68.2, capability 65.9. The v1.5.4 page opened September 30 keeps 68.2 and lists official rank 13. The proxy is not that composite.
swift-jev types Choice, Noul, and Score, and routes at 0.60 and 0.85
Daiki Matsudate posted a Swift client after a trial that paired Jev with another model did not go well. The library is MIT, has no dependencies, and returns enum cases, probabilities, and scores from one request. The sample policy escalates under 0.60 and acts at 0.85 or above. The README prints no scored set.
djev reads DiffusionGemma as a decision API, and sits third on JevBench
Florian S's JevBench table puts djev at 74.3, 1.1 behind Jev. Maisa's djev.dev hosts the API at an announced $0.035 per million input, with native image questions. Matt Mastracci posted phone-camera clips of DiffusionGemma-as-Jev. Davipar/djev-dev is the Apache 2.0 local stack.
A native DuckDB extension batches Jev over SQL rows at about 2,300 rows per second on a live run
Prasanth J posted duckdb-jev, an Apache 2.0 C++ extension that asks TypeSafe's Jev from SQL. A live Choice run classified 2,049 synthetic rows in 0.887 seconds. A cached replay of the same query finished in 28.6 ms with zero API calls. Row text leaves DuckDB.
jevsearch re-ranks keyword hits with Jev and scored 83% Hit@1 on the author's TypeSafe docs set
Kyle McLaren posted jevsearch, an MIT shadcn registry block that streams keyword hits and then re-ranks them with Jev. On 41 labelled queries over TypeSafe's documentation, Hit@1 rose from 41% on the keyword pass to 83% after Jev. Median uncached ranking was 278 ms. About $0.26 per thousand searches on that corpus.
Open-Jev 27B v1.1 answers 197 of 231 public JevBench tasks
Zefan Cai released Open-Jev 27B v1.1, a LoRA and decision head on Qwen3.8-27B. On the 231 public JevBench tasks it is correct on 197, three behind Jev 1.13.0 at 200. The hard tier is 80 of 111 against Jev at 81. The September 21 audit of the released 2B and 9B checkpoints, 150 and 179, is unchanged. The card prints no speed comparison. On October 2 one 64-step pilot on the released 2B checkpoint moved a 128-row synthetic test from 84 of 128 to 99 of 128. Those rows are not JevBench.
SimpleJev turns an open model into a classifier by reading the next-token logits
Featherless posted SimpleJev, an Apache 2.0 server that scores Choice, Score, and Noul from a Hugging Face model's next-token logits. A public demo needs no key. JevBench listed Qwen3.8-27B at 67.3, with hard-tier accuracy 75.0%, matching Jev on that slice and trailing on the composite.
Spring publishes a Java client that asks Jev before the chat model runs
Opened October 8, 2026, the README quick start pins 0.4.0, and Maven Central lists that release. Christian Tzolov wrote up spring-ai-community/spring-ai-typesafe, an Apache 2.0 Java SDK and a set of Spring AI advisors. On his laptop a three-question ticket call took 310 ms. The blog says 0.1.0 is on Maven Central. The README quick start still shows 0.1.0-SNAPSHOT, and a search of search.maven.org returned no documents for those artifact ids. GitHub release 0.3.0, September 30, says the same client talks to Laya. The Ollama page is on the snapshot docs. A Maven search on October 1 still returned no documents for spring-ai-starter-typesafe.
DocJev classifies and splits real PDFs with Jev, 5.7x to 6.5x faster than Luna on the author's set
Jerry Liu posted DocJev, an Apache 2.0 library that classifies and splits PDFs with local LiteParse and hosted Jev. On 40 public-sector originals, both engines scored 40/40. Jev split 7 of 8 packets exactly against Luna 8/8, at 5.73x and 6.45x the median decision speed.
Jared Palmer's Kev family copies Jev's API on Qwen, and trails it on new sources
Cognition's VP of engineering published Kev, an Apache 2.0 family of Jev-like models you train and serve yourself. On data the models never trained on, Kev-8B scored 79.6% against Jev 85.7%. JevBench listed kev 0.6B at 66.7. A Qwen3.5 generation followed the same day. On September 25 he posted Kev-4B on OpenRouter. A September 26 paired run scored that checkpoint at 1,216 of 1,253 full-schema function selections, against 1,235 for a fresh jev-1.13.0. On September 28 Christoph Görn scored the same 4B at 1 of 7 on eight of his own prompts. On October 2 Jacob Lin's StreamDecisionBench chart prints Kev-27B at 62.15 and the Jev API at 59.63, as normalized log-AUC over 0.5 to 8 seconds.
Laya, an Apache 2.0 System One stack, publishes where it beats Jev's numbers and where it does not
Louis Grenard pointed at Laya, Convai Innovations' open typed-decision engine. Three checkpoints, a Router that picks among them, Apache 2.0 weights on Hugging Face. The author's table puts routed Laya above published Jev on some English sets and far below on Banking77. Jev was not called. JevBench's independent row is 70.1 against Jev 75.3.
semgate asks Jev about each HTTP request, then a live range to attack it
mizutani published semgate, a Go net/http middleware that sends each request to Jev as typed questions and lets your function block, route, or pass. A companion range, Injection Range, is up on Cloud Run. Default configuration sends every header, including Authorization and Cookie, unless you denylist them.
GitHub Next ships LocalJev, a prompted stand-in that speaks /v1/systemone
GitHub Next published LocalJev, a Bun server that turns Jev questions into JSON-probability prompts for models running on oMLX. A 1,200-request bake-off on an M5 Max compared five 4-bit checkpoints. The authors say the probabilities are self-reported, not read from logits.
LlamaIndex's Logan Markewich ships jeff, a local System One API on GLiFormer
Logan Markewich released jeff, a self-hosted drop-in for TypeSafe's /v1/systemone endpoint backed by Knowledgator's GLiFormer. On 1,600 public items it trails Jev, especially on irony and reading comprehension, and comes in cheaper to serve on a GPU.
Jevis is a Flutter integration-test agent that only taps what you registered
Jake Gwon released jevis, a Dart package that drives Flutter's integration_test loop with Jev. You register legal actions, write a goal, and set an action budget. Jev picks the next widget action and a Noul checks whether the goal is already on screen.
SemIf, a 4B open rebuild that reads option logits, sits 0.7 behind Jev on JevBench
Theodore Lee's SemIf (formerly OpenJev) scores typed options from a frozen Qwen3.5-4B by reading logits, not by sampling JSON. On JevBench v1.2.2 it is 74.6 against Jev 1.13.0 at 75.3. The author's own 3090 timings put direct readout at 1.023 s for 21 questions against 5.332 s for a compact JSON array.
Bespoke Nimble 9B matches 90% of its own labels and 74.8% of a public suite
Bespoke Nimble 9B scores 90.12% vs 93.21% for Jev 1.13.0 on 324 synthetic holdout labels. On 13 public subsets the repository macro is 74.8% against Jev at 76.0%. Ollama's later mean on the same 3,880 records is 75.7% for Nimble, with Jev still listed at 76.0%. The September 24 checkpoint lengthens the context and prints no new score. llama.cpp pull 29844 names Bespoke-Nimble-9B-v3. That card prints Decision Index 0.2.1 at 56.88 and CC BY-NC 4.0. It is not the 324-row holdout.
Hono's author routes HTTP by meaning, then warns you not to
Yusuke Wada published hono-jev-router, an experimental Hono router that asks Jev which plain-language description matches an incoming request. The README treats it as a demo with sharp limits.
A week of Jev repos: routers, MCP, graphs, games
A widely shared Awesome Jev thread listed fourteen working patterns; community directories passed a hundred projects within days. The repeating pattern is choose, score, and stop.
Zachi's pg-jev judges 2,000 Postgres rows in about 3.5 seconds
Zachi's pg-jev judges a 2,000-row Postgres table in about 3.5 seconds for about $0.012. Version 0.2.0 packs 20 rows per request. The page at pgjev.com still says 40. Jose Mejias's pg-jev is a different repository.