Dated, sourced news in the System One / decision model category: launches, integrations, benchmarks and adoption, linked to the model it affects.
On the TAB test split (127 court judgments), Clairic + Jev keeps 96% recall on names and case IDs and lifts precision from 78% to 91%, at $0.86 per 1,000 judgments.
alphaXiv and OpenResearch release an Apache-2.0 rank-16 LoRA for typed decisions. Author-run JevBench public set: 83.55% vs 76.19% for the base. No independent board yet.
An X article walks through pplx-decider v1.2: non-causal full attention, a 255-way readout at a mask token, tasksource-style training and a fitted temperature of 1.32.
Drex Max v1 weights are out under Nace's Open RAIL-M, with Q8 GGUF and a Q4 for 24 GB Macs. Decision Index 0.4.1 Full 66.60 (#24); the card's other numbers are Nace's own.
Base Labs adds a 2B model to the mammoth-1 / elephant-1 family: MIT weights, Baseten API at $0.01/M input. Strong public split, but Decision Index Full is 35.91 (#148).
Braintrust ran Jev 1.14, D1 Ultra, GPT-6 Luna, D1 and Clef on the AID test set. Accuracy 83.85–86.15%; only Jev vs Clef is significant. Jev best calibrated.
PelaAI's KnowLine-31B-Gen1 (Apache-2.0, Gemma 4 31B) scores Full 71.86 on Decision Index 0.4.1, ahead of GLiDE 1.2. Turion Jev 31B lands at #4 with 70.47.
Second Lumma-Fev generation: Apache-2.0 4B and 27B, trained on ~300K samples with decisionformer. Vendor numbers beat Jev 1.13 on eight benchmarks; the board PR is still pending.
Fireworks now serves Rune 26B-A4B (surogate/rune-26b-a4b) serverless and on demand: 262K context, text only, output free. The model's first hosted API from a large inference provider.
Onur Solmaz's MIT TypeScript framework: your program offers a short list of next steps and a decision model such as Jev picks one at a time. First domains are chess and eval judging.
23 methods on 72 synthetic small-business bank statements: Claude Opus 5.5 leads out of the box (95.5%); a Qwen3.5-4B decision model tuned on 24,400 labels hits 97.3% at ~1/80 of the cost.
Nace.AI's Drex Max 1.1 is a hosted decision model with images and a 262K context. Nace reports 70.8 on the Decision Index; the official board has no 1.1 row yet.
Sardine CEO Soups Ranjan says Jev reached 93% recall on a real fraud incident and is being added to Sardine's platform. Panel with TypeSafe CTO Erik Gafni in SF on 4 Nov.
Sthanika AI's Sieve-31B-v3 (Apache-2.0, Gemma 4 31B) scores Full 71.10 on the Decision Index 0.4.1, #2 behind GLiDE 1.2 and ahead of d3-ultra.
Vela 3.0 (Apache-2.0) runs decisions, router signals and span extraction in one call. The 0.4B vela3-fast scores 50.36 public but 29.27 Full on Decision Index 0.4.1.
d3-ultra (Apache-2.0, 26B, text, images, video) is #2 on the Decision Index 0.4.1 at Full 70.18, behind GLiDE 1.2. The d3 family is now on Ollama 0.40.11 and Tinfoil.
LFM3 ships six open ~2B models; d1-2B (text, images, video) and d1-Omni-2B (adds audio) are decision models with typed answers in one pass. Not on public boards yet.
Razer's local AI toolkit adds Mapika Decider and AutoTrust Jev. In Razer's own DOOM test, Mapika Decider 0.8B (40 ms) beat Jev over the cloud (270 ms). Vendor-run.
AI Engineer's Code Summit in San Francisco (10–12 Nov 2026) lists a 'System One Models' keynote by Diogo Almeida, co-founder and CEO of TypeSafe AI.
Wan et al. score topics, then documents, then passages with a decision model, one pass per level. They report +13.8 R@1 on MultiDoc2Dial and +17.4 Hit@1 on NQ over dense retrieval plus an 8B reranker.
Singularity Compute's Apache-2.0 Super Decider scores 78.5 on JevBench's hosted-API board, ahead of kettei-fat (77.7) and GLiDE (76.4). Hosted on ASI:Cloud at $0.03/M input.
Alyona Vert's guide groups 15 research decision models by approach: candidate scoring, shared encoding, video/audio/spatial, bi-encoders, loops, hybrids and distillation.
FrontiersMindAI's Apache-2.0 library: typed choice / noul / score in one forward pass, one-command fine-tuning with QLoRA, /v1/systemone server and a vLLM backend. A runtime, not a model.
StartLux Labs' inference engine runs the released GGUF files as int8 or W4A8 behind /v1/systemone; author claims 19 ms vs 53 ms for vLLM FP8 on one H200 and bit-identical answers.
arXiv 2610.29961 adds a none-of-these condition to CaseHOLD. pplx-decider matches Gemini 3.8 Flash at ~1/55 the cost; Jev picks a wrong holding least but over-abstains.
Menu-bar app for Apple silicon: pplx-decider-flash-v1.2 as a 5.8 GB q8_0 GGUF at 127.0.0.1:8080/v1/decisions, no API key, text and images, one question per request.
Passau/IT:U study on six Qwen backbones: reranker fine-tunes beat decision fine-tunes by 5–29 nDCG@10 up to 9B and tie at 27B; calibration gains depend on the question wording.
TypeSafe's CEO posted a mock breach notice on 23 Oct; the linked gist holds the 101 possible noul values. Some outlets reported it as a real incident.
GLiDE v1.2: 27B on Qwen3.8, weights on Hugging Face under a revenue-capped community license. Decision Index 0.4.1 #1 (71.52, no tie); JevBench API Intelligence 76.4.
GA on 22 Oct 2026: still gpt-6-luna only, $0.10/M input ($0.01 cached), JSON input, images per question, GPT-Live integration, US/EU residency. Stays a runtime.
Subquadratic's Snap 1.1 scores 65 on AA's Decision Index v1.0, level with d1-ultra and Elephant 9B, 8 points over Jev 1.14. Price drops to $0.04/M input.
The 27B and 9B decision models that were gated on 20 Oct are now MIT, with a /v1/systemone API on Baseten and Vercel AI Gateway. Decision Index 0.4.1 rescored the public mammoth-1 at 68.44.
REAX (decisionmodels.io, Bittensor subnet 92) released REAX 1 under Apache-2.0. The only numbers are its own. JevBench's operator, who also runs REAX, says it will not be listed there.
JevBench v1.6.32 (20 Oct) ranks Chon-1.0-12B first among Jev-class open-weight systems: Capability 85.3, ahead of basal-omni-1.0-ultra and Quyet-1.0-Large.
On 100 labelled agent traces, OpenAI Decisions led on F1 (0.87) at about 4.5× Perplexity's cost; Jev and d1-ultra were already well calibrated. Advice: don't keep the 0.50 default.
Decision Index 0.4 (20 Oct) puts baselabs' mammoth-1 first: Full 71.24, clear of the tie band. Its weights are gated for Decision Index evaluation only, so it is not in our catalog.
UiPath Integration Service now has a TypeSafe connector in preview. It calls Jev with the customer's own TypeSafe API key and returns a typed answer, per-option probabilities and confidence.
Clairic (preview, Workers AI, $0.10/M input) extracts, computes and forecasts typed values with confidence ranges. Open weights promised soon. Vendor numbers unverified.
d3-turbo (Apache-2.0, 26B) answers every question from one read of the input. Vendor says #1 on both Decision Index boards; DI 0.4 puts it at Full 69.04 (#6), level with d3.
Vals AI asked Jev and Mercury Decide questions such as 'Is Taiwan a part of China?' and reports answers leaning more pro-China than most frontier models. A small probe, no published method.
Zeki Works ran 24 decision models on 2,198 Turkish questions; WaldoBench gave 11 image models 66 Where's Waldo scenes. d1-ultra led both. Jev 1.14 was 59.2% on the first and fastest on the second.
Nandakishor M's Vega-U shares one 14.3M decision engine across frozen Qwen3.5 0.8B and 4B, Qwen3.8 27B and Gemma 3 1B readers. Accuracy and calibration are the author's; Apache-2.0; no board.
Upstream vLLM v0.32.0 (18 Oct) serves choice, noul and score questions on /v1/systemone by default; transformers v5.20.0 has the endpoint in `transformers serve`. Not the third-party vllm-jev.
AutoTrust's 27B text decision model v1.1 (Apache-2.0) reports 67.31 on the DI 0.3 public suite, its own run, pending maintainer verification. Not on DI 0.4 or JevBench.
The JevBench maintainer says 34 models were added or newly ranked since Thursday. Board check: kettei-fat is #1 on API (77.8) and ImageJevBench (75.39); Sage 1.4, Celeris-1 and basal-omni are in.
TypeSafe's Jev 1.14.0 is on Databricks Unity Gateway (16 Oct), hosted by TypeSafe, behind an admin-enabled preview. Databricks also lists OpenJev (Qwen3.5 4B) as a pay-per-token service.
Netlify's AI Gateway serves TypeSafe's jev-1.14.0 to Netlify Functions through the official SDK, with credentials handled by the gateway and usage billed to Netlify credits.
Perplexity released pplx-decider-v1.2 (27B) and Flash v1.2 (4B) with Apache-2.0 weights. The 27B is #4 on the official Decision Index 0.3.2 (Full 65.55); API from $0.01/M input.
StartLux Labs' v1.1 models (Instant and AutoThink, 0.8B-27B, CC BY-NC) are on the official board: the 27B scores Full 69.12 on Decision Index 0.3.2, first and about 3 points clear of d3.
A new Atlassian Automation step uses TypeSafe's Jev to check, categorize and score text, with probabilities and confidence returned as smart values. Early access is open; pricing is not set.
LLM Gateway's coding agent calls its /v1/decisions API before each turn to choose Haiku, Sonnet or Opus. Jev 1.13 is the default; Microsoft-Decision-1 and GPT-6 Luna Decisions are selectable.
UIUC, Princeton and Google researchers built AgentDecisionBench and a harness that hands tool choices to a small decision model when it is confident. A 4B trained on agent states beat Jev.
Liquid's larger hosted decision model reads text and images, costs $0.042/M input and scores 65 on AA's Decision Index v1.0, mostly by abstaining when it doesn't know.
HiddenLayer (14 Oct) steered an agent harness with an ad; Datadog (15 Oct) cut Jev's injection recall from 82.90% to 3.88% with a prepended fake policy. Both tested flagging, not whole systems.
raxIT's Decision Models as Guardrails v1.0.0 scores 19 systems on 9,975 checks. Perplexity 90.4, OpenAI 90.0, Clef 89.5, Jev 1.14.0 88.7, Bedrock Guardrails 79.0. Numbers as reported; part is private.
Span-02 is a Gemma 4 26B-A4B fine-tune with its own engine, a number answer type, images and per-request thinking budgets. Weights on Hugging Face; $0.01/M input hosted.
MachineFi's hosted decision API now serves /v1/systemone and reports 90.9% and ECE 0.019 on a ten-task panel it trained on. Vendor-run; JevBench submission queued. We re-scored the page.
A free System One API serves six Decision 3.0 models across text, image and video; a key is available without an account.
Decision Models serves existing open decision checkpoints through REAX; its founder also discloses that he runs JevBench.
Image input is the standout: pictures can go anywhere in a request. Jev 1.13 already had it, and images, box and 262k context are “danger zone” opt-ins, not production-ready. Numbers are internal.
Sqwish Labs launched OneHot with kettei fat, chub, core and dot. kettei-fat is #1 on the JevBench API board (77.8) and ImageJevBench; core and dot are called open source, but no weights were found.
Zaid Mukaddam (Scira) releases open weights plus a free API. Author run: DI 0.3 public 65.11 for the 27B, which would be #3 on the public column.
Independent index of 11 decision models over four sets. Jev 1.13 58, GLiDE and Microsoft-Decision-1 54. Decision models are faster and cheaper than LLMs but overconfident.
Closed hosted models at $0.04 and $0.09 per million input tokens. Claims +6 and +10 over Jev 1.13 on DI public scores; not on the board.
JevEval scores a test case as the weighted mean of bounded questions answered by Jev. Eval modes also let built-in metrics run as llm, hybrid or system_one.
Drex Max (hosted only, $0.06/M input) reads text and images on /v1/systemone. Nace reports DI 0.3.2 public 67.59; DI 0.4 lists it at Full 68.51 (#8). Vendor numbers otherwise unverified.
NUS researchers encode the request and each option separately, then refine only the request state. Jeva-4B shows zero order flips and strong null detection in the authors' tests.
Tokyo's Jurin AI releases a four-size family built on StartLux-Decision. The 79.20% headline is the private 9B; the open 2B scores 68.54% on the same vendor report.
arXiv 2610.13092 tests Jev in five agents over 4,500 runs on ALFWorld and AndroidWorld. Cost and latency drop; task success holds or falls, more in mobile agents.
Drex 2.0 is a Qwen3.6-35B-A3B decision model (~3B active) with a pointer head. Nace's public Decision Index 0.3.1 run is 61.25. Hosted price is $0.06/M input.
Closed, multimodal, built on sparse attention. Claims DI 0.2.1 61.7 vs Jev 57.9 and under one second at 100K tokens. Public preview with limited capacity.
Independent quantizations of pplx-decider 27B, Clef, Clef-Flash, Kev and d3. The 27B drops from 53.8 GB to 15.3 GB and matches the original on 96.8% of 2,000 decisions (vendor).
Bailian's decision-model-preview speaks POST /compatible-mode/v1/systemone. The docs do not say the model is Jev. Beijing and Singapore list it as free for a limited time.
The 0.3.1 community vision board puts JEV-27B-VL second of 21. Its 77.8% accuracy claim matches the board data; the post's first-place claim is stale.
Apache-2.0 MoE decision model, up to 256 options per pass. Own Decision Index 0.3 public-suite run: 59.75 (NVFP4 59.18). 241 ms median on a B200.
EdgeEvals credits pplx-decider v1.1 on 643 of 669 authored cases. It shares rank 1, with one severe error; the preview is not clinical validation.
FluidInference publishes an Apache-2.0 Qwen3-0.6B student distilled from Clef-flash, with typed outputs and a Core ML runtime for Apple silicon.
d3 27B to d3-edge 0.6B, Apache-2.0. The 27B claims 65.45 on the Decision Index 0.3 public suite (own run, official kit), +8.5 over Decision 2.0.
LiteLLM v1.104.2 adds /v1/decisions and /v1/systemone routes for typed-decision providers; its Playground can test configured models.
Benchmark Heaven measures Microsoft-Decision-1 on 1,500 decisions: sixth among hosted APIs, with strong category results and a lower composite than Jev.
THX-01 is now available through HAL-X's hosted /v1/systemone endpoint. The announcement says free and keyless; current API docs require a HAL-X key.
ZekiWorks ran 2,198 non-English exam questions five times. Jev changed 314 answers, Liquid d1 217, pplx-decider v1.1 and Open-Jev none.
Clef-omni (Qwen3-Omni-30B-A3B, Apache-2.0) scores decisions over text, images, audio and video at $0.15/M. Clef-flash now costs less than Jev but its hosted window drops to 24K.
Drex v1.5 (MiMo-Qwen 9B + pointer head) scores 50.84 on the official Decision Index 0.3, public part 58.08. Free only for organisations under $1M revenue.
Apache-2.0 LoRA + head checkpoints from 2B to 26B-A4B with a /v1/systemone server. Strands’ own public-split numbers: 77.9–89.6%. Not on any board yet.
Andreessen Horowitz leads, with Sequoia, DCVC and angels; Martin Casado joins the board. TypeSafe says a third of the Fortune 500 use Jev; a16z says 25% have integrated it.
The Hugging Face Decision Index 0.3 now weights private tests 80%. pplx-decider v1.1 leads at 62.75 ahead of Jev (60.11); Perplexity cut the API to $0.02/M.
Benchmark Heaven re-ran every system on a fresh v1.6 item draw. Scores are not comparable with v1.5. Quyet-1.0-Large leads open weights; Sage leads the API board.
Liquid AI published two open decision models on 7 Oct: d1-3B (text and image, 3.1B) and experimental d1-omni-600M (image or audio). They are not the hosted d1 API.
Unsloth's open repo can fine-tune Qwen, Gemma and Llama with a Clef head and LoRA, then serve the result on a local Decision API. No checkpoint was published.
POST /v1/decisions with gpt-6-luna: predicates, choices and scores over text and images, input-only pricing. First independent result: JevBench API board #5 (62.5).
111 ranked systems. Official score: Cygnet 73.7, Winnow-12B Q8 73.2, Jev 1.13.0 72.1. New capability ranking puts Jev at 80.0. Numbers verified from the live board.
Brazilian developer Fabio Akita published a long Jev explainer with an audit by Paulo Câmara. We re-checked its four recomputed studies from the raw data. Jev's capability goes from 8.2 to 8.1.
The Swarms agent framework's v16 "Overclock" release ships DecisionModel: one interface for TypeSafe's Jev and Cloudflare's Clef, plus a /v1/systemone route in the Swarms API.
Same weights as JEV-Gemma4-26B-A4B under a new name, plus opt-in thinking and computer-use and robot demos. AutoTrust's own Decision Index score of 62.48 is not a board result.
Ollama serves Clef and Clef-flash on /v1/systemone, ggml-org GGUFs wait on an open llama.cpp PR, and an independent 85-case test finds Clef accurate but slower than Jev on hosted APIs.
Jev files uploads into folders and walks the tree to find evidence, with no vector database. The code is public, but the repo has no license file yet.
A T5-encoder decision model under MIT with an int8 CPU build and a vLLM plugin. Its 0.893 vs Jev 0.897 is on razvilka, a benchmark the same team published that day.
llama-server now answers System One requests for Laya, Julia-1, Lev, OpenJev 27B and Kev, with official GGUFs for all five. Clef text support was merged the same night.
Nine guardrails on two English tasks. A 200M classifier and a 35B judge lead prompt injection; Jev leads content safety. Red Hat's verdict: decision models don't reliably win.
Diogo Almeida told the Wall Street Journal that ~25% of Fortune 500 companies use Jev and traffic hit a trillion tokens a day. Both figures are his alone.
Six Apache-2.0 sizes, Kai-0.6B to Vega-27B. The team reports Vega at 56.5 on the Decision Index (Jev 57.91) in its own run with the official kit; not on the public board yet.
Clef (27B) and Clef-flash (9B) ship as Apache-2.0 weights and as Workers AI endpoints at $0.24 and $0.09 per M input. The 'leader on the Decision Index' claim is Cloudflare's own run.
GLiDE gives a fast distribution, then reasons when unsure. Hosted only, $0.30/M input. Fastino reports 64.81 on Decision Index 0.2.1 in its own run; GLiDE is not on the public board.
InfoQ recaps Jev's pricing and latency and collects adoption and third-party numbers: Vercel's 24-hour uptake, and an OpenChamber analysis that puts the median speedup at 7×, not 193.6×.
Four Apache-2.0 models from 0.8B to 27B, TypeSafe SDK compatible, with GPU and MLX serving and fine-tuning skills. New Kev-27B is a full SFT of Qwen3.8-27B. Numbers are the author's.
Two days after launch, d1 is on OpenRouter as liquid/d1: $0.04/M input, $0 output, 65K context. OpenRouter's post says zero data retention. First third-party distribution for the model.
pplx-decider-v1-27b costs $0.04/M input. Its Hugging Face weights are byte-identical to AutoJev-27B, already in the catalog, so we updated that page instead of adding a model.
A Qwen3.5-2B with a pointer head and LoRA, Apache-2.0 with data and scripts, served on /v1/systemone. JevBench 0.723 is AWS's own run.
bandr.ai's recipe LoRA-tunes Qwen3.5-4B on labelled agent steps. Its own run beats Jev on AgentProcessBench, in-distribution. No checkpoint or license yet.
Surogate Rune, AutoJev-27B, Mica 4B, NeoHorse-Jev-4B and AutoTrust JEV were trending in text classification on HF. Two sit near Jev on the official Decision Index; one is a Jev distillation.
Australia's Maincode publishes Apache-2.0 Jev-style weights with choice / noul / score, images and a local /v1/systemone server. The Index is Maincode's own run and isn't on the official board yet.
Bespoke-Nimble-9B-v3 is a new LoRA on Qwen3.5-9B for typed choice and yes/no. Bespoke's own DI 0.2.1 run: 56.88. The license moves to CC BY-NC-4.0; V1, V2 and Ollama's nimble stay Apache-2.0.
Ollama 0.35 serves decision models locally behind TypeSafe's API shape. Launch models: Nimble 9B (Bespoke Labs) and Tev1 4B / 0.8B (Together AI).
Trio-Spark v1.0 is a hosted Choice decision API from MachineFi, submitted by an affiliated party. Listed as beta; lineage undisclosed, all numbers vendor-reported.
Liquid AI's first decision model ships as a TypeSafe-compatible API with a free tier. The Decision Index win (58.9 vs 57.9) is Liquid's own run; d1 isn't on the official index.
At DevDay, OpenAI announced a Luna-powered API that picks from answers you define. Limited preview; no docs, schema or pricing yet. Listed under runtimes, not the catalog.
Qwen3.5-9B with a LoRA and a pointer head, trained with SFT and CISPO, served on /v1/systemone. Claims 0.935 on JevBench's public items; thinking costs a 3.3 s median.
StartLux Labs publishes open decision models from 0.8B to 27B (plus a 35B-A3B MoE on 1 Oct) and GGUFs. Weights are CC BY-NC-4.0. Author tables put the 27B above Jev; not on the official board.
The type-safe generation runtime now has a hosted service: playground open to all, API rolling out to early-access users, $0.05/M input and $0.50/M thinking tokens.
Apache-2.0 LoRA on Qwen3.5-4B adds photos to /v1/systemone, with a trained can't-tell. The board ranks are cited by the author (UnverifiedClaim). It joins the catalog late.
HIT-TMG's embedding-to-decision stack ships JevEmbed-Qwen3-Embedding-4B (LoRA merged). Author in-domain acc 85.86% on JevEmbed-Data = UnverifiedClaim. Repo ~47★.
The local decision-model runtime gains Mapika's decider:2b-vision and an images field on the API. Author RTX 4090 ~200 ms for five questions on a small frame = UnverifiedClaim. ~786★.
Shanghua Gao ships RSI-Jev v3.0 (Qwen3.5-2B): data-scale-2 SFT plus listwise ranking RL. Suite 0.756 / typed 0.791 / ECE 0.066 = UnverifiedClaim. Local /v1/systemone.
Damien Sileo / tasksource releases an Apache-2.0 LateOn fine-tune for choice/noul/score via ColBERT late interaction. Author benches (Typed 42.6%) = UnverifiedClaim.
chaoliangUNSW releases Jev-Style-2B-Decision-v3 (plus GGUF/MLX): same v3 recipe at 1.88B, jev-style≥0.3.0 --release 2b. Self-run JevBench 73.6% = UnverifiedClaim.
OmniJev releases four Apache-2.0 decision models that answer Noul / Choice / Score about text, images and video, plus the training data. All benchmark numbers are the authors' own.
manjunathshiva releases OpenDecider nano/small/medium-td with pip install, Colab and a Space. Author typed-decisions 0.796 vs Laya/Jev = UnverifiedClaim (nano in-domain).
Shanghua Gao releases RSI-Jev v2.1: Apache-2.0 Qwen3.5-2B fine-tune with a trained scoring head, confidence calibration, and a local /v1/systemone server. Author suite 0.7291 = UnverifiedClaim.
A widely read X article by @polydao shows three Claude Code hooks that hand yes-or-no calls to Jev and lists seven rules for writing typed questions. The cost math is the author's.
David Zhang released jevgrep, a CLI that uses Jev to find relevant code for coding agents, claiming 40% cost reduction on SWE-bench.
Brazilian lab Supersonic Labs opens Julia-1: Apache-2.0, fine-tuned from mmBERT-small, choice/score/noul in one API. Author evals put it level with Jev on typed decisions; all UnverifiedClaim.
Jared Palmer's Kev 4B is now available on OpenRouter as jaredpalmer/kev-4b — the first open-weight System One model on a public LLM gateway.
OpenRouter ships typesafe/jev-router, a cache-aware model router that uses Jev to pick the best model for each request.
The pi-jev v0.6.0 release adds custom Jev-compatible endpoints, endpoint status, OpenCode Zen routing, and runtime fixes for the Pi coding agent.
Respan's Span-01 outscores Jev on behavior detection with an open benchmark — and then uses itself as the eval layer for 11 catalogued decision models, publishing their ECE.
mode-io's vllm-jev serves Open-Jev, Laya, Tiny-Jev and Valen over /v1/systemone with vLLM on Linux and MLX on Apple Silicon. ~94★; speedups = UnverifiedClaim.
Caio Vicentino published Eikos-4B and Eikos-27B (MIT, Qwen fine-tunes) with FP8/INT4/MLX builds, the training set and the full eval harness. Benchmarks stay UnverifiedClaim.
A new open-source runtime lets you serve Laya, Decider, Kev and other decision models locally behind a TypeSafe-compatible API — desktop apps, CLI, Docker, GPU or CPU.
Valen-Team ships Valen-Preview-0923: Qwen3.5-2B + decision head (SFT + experimental RLCD) for Choice/Noul/Score over text, images and video. Author Sokoban benches = UnverifiedClaim.
Official @googlegemma amplifies taeold/djev-run — one-command Cloud Run deploy of a Jev-compatible DiffusionGemma-Jev endpoint. Latency and pricing claims stay UnverifiedClaim.
FLock.io releases MIT this-that-model-1.0 (~1.88B hybrid Qwen3.5/DeltaNet) — declared options, calibrated probs, zero output tokens; self-pub benches stay UnverifiedClaim.
Independent Apache-2.0 Jev-Omni (Gemma 4 12B) ships on Hugging Face — four-modality option classifier; author benches and launch-video latencies stay UnverifiedClaim.
Community Apache-2.0 demo runs local Laya (MLX) beside hosted TypeSafe Jev on Chrome’s T-Rex course, with live metrics and replay — assisted scores include safety shields.
Packagist lists symfony/ai-type-safe-platform — official TypeSafe bridge for Symfony AI (PHP ≥8.2), published 21 Sep 2026.
Independent Apache-2.0 Tiny-Jev (Qwen3-0.6B LoRA merged) ships on Hugging Face — calibrated typed decisions, no text decode; author benches stay UnverifiedClaim.
@ItsCuthulhu runs benchmark of 40+ Jev alternatives. Partial rankings: djev (#1), simplejev-qwen38-27b (#2), Reflex-4b (#3), Decider-2b (#4), Laya (#5). Live bench at bench.jakecuth.com.
LangChain blog (20 Sep) compares Jev to LLM judges on fixed Deep Agents traces — lower variance, cost, and latency in a narrow LangSmith experiment; separate from the harness post.
YouTube walkthrough of SemIf, Nimble, Decider, openjev, DiffusionGemma, NanoJev, and Laya — plus JevBench. Catalog gap filled: AlexWortega/openjev.
Bespoke Labs' open Jev adapter can now process images; demo shows floor-plan counting where base Qwen failed.
trycua’s article explains hosted Jev with Cua Driver (jev-use) and an independent form specialist (CUA-S1-FORMS) inspired by jevlike — not a Jev fine-tune.
Ksenia Se’s skeptical deep dive on Jev and RLCD — research lineage, open alternatives, and why confidence ≠ probability.
Context7 compared Jev 1.13.0 to Gemini Flash and DeepSeek on five classification tasks used in their parsing pipeline.
Official @googlegemma posts on DiffusionGemma running structured decisions in one parallel pass — awareness for djev-spark, not a new TypeSafe model.
Grep.ai CTO Miguel Ríos announces AgentRun, a workflow harness that pairs frontier agents with TypeSafe Jev typed questions for repetitive regulated knowledge work.
TechCrunch, InfoWorld, and The Register cover TypeSafe's Jev as typed probabilistic decisions for software—not chat.
Jev runs on Vercel AI Gateway and OpenRouter. Vercel's October index: 35.3% of paid teams in 14 days, a gateway record, but only ~0.3% of spend. The 13% day-one claim stays unverified.
Official LangChain post introduces langchain-typesafe with TypeSafeClassifier, ModelRouterMiddleware, and AutoModeMiddleware.
Demand briefly knocked the Jev API offline as the launch discussion hit 1,890 HN points.
TypeSafe AI exits stealth with ~$40M seed and introduces Jev — typed decisions, not chat.
