Jev directory
CommunityJev Models & Reimplementations
PurpleMIST-Flash-1.0
OlaVerse's Apache-2.0 System One decision model from a stripped Qwen3.5-9B-Base: it returns calibrated choice, noul and score distributions for English and eight African languages in one pass, up to 65,536 tokens.
- Reading time
- 3 min
- Published by
- Community
- Author
- olaverse
- Last verified
- Oct 2026
- Languages
- Python
- Capabilities
- Open sourceSelf-hostedRuns locally
Highlights
- 7.9B model built from Qwen3.5-9B-Base with the vision tower and LM head removed: 24 Gated DeltaNet plus 8 full-attention layers, 15.9 GB in bf16.
- 0.668 accuracy on African Typed Decisions against 0.669 for the 1.6T DeepSeek-V4-Pro, with KL 0.214 vs 1.125 and Brier 0.115 vs 0.199.
- Covers eight African languages plus English; accuracy drops 0.053 from the English original to the translated set where OpenDecider-small drops 0.171.
- Answered all 140,620 Jev Decision Index requests, some over 65,000 tokens, without truncation; Decider defaults to 2,048 tokens to keep memory low.
- serve.py exposes POST /v1/systemone matching Ollama's decision API, but PurpleMIST is not in the Ollama library yet.
Watch out
Apache-2.0, 15.9 GB of bf16 safetensors needing a 24 GB GPU, in a repo with 73 downloads. The evaluation table is the card's own harness: LLM baselines were prompted for JSON probabilities, DeepSeek-V4-Pro is scored partly against its own native-case answers, and the native column marks PurpleMIST as trained on cases from the same pipeline. The local server handles one request at a time.
Questions about this entry
Short answers drawn from the linked project or page.
A 7.9B System One decision model from OlaVerse built on the text half of Qwen3.5-9B-Base. Given unstructured state and typed questions it returns a calibrated probability distribution for every answer in one forward pass; it scores options and does not generate text.
Install transformers>=5.19, torch and huggingface_hub, then load it through Decider.from_pretrained. For an HTTP endpoint, download serve.py and purplemist.py and run python purplemist/serve.py --port 8000; the server offers POST /v1/systemone, GET /health and GET /v1/models.
English plus eight African languages: Yorùbá, Hausa, Igbo, Nigerian Pidgin, Kiswahili, Amharic, Somali and isiZulu, matching the eight-language African Typed Decisions set. Instructions and option descriptions can be written in English or in the state's language.
Up to 65,536 tokens per pass, and the card says it answered all 140,620 Jev Decision Index requests, including ones over 65,000 tokens, without truncation. Questions that do not fit alongside the state are split across several passes automatically; Decider defaults to 2,048 tokens and accepts max_len=65536.
On the card's African Typed Decisions run it scored 0.668 accuracy to DeepSeek-V4-Pro's 0.669 at roughly 200x the size, with its distributions about 5x closer to the reference (KL 0.214 vs 1.125) and Brier 0.115 vs 0.199. It leads on Pidgin, Somali, Swahili and isiZulu and trails on Hausa and Igbo.
Not yet. serve.py returns the same answer fields as Ollama's 0.35+ decision endpoint at POST /v1/systemone (choice, noul, score, probabilities, confidence, legend), so clients written for it work, but the card says PurpleMIST is not in the Ollama library and should be run with serve.py for now.
Apache-2.0 per the model card front matter, covering 15.9 GB of bf16 safetensors with the unmerged LoRA adapter in the adapter/ folder. A smaller 1.9B sibling, PurpleMIST-Mini-1.0, shares the interface for English workloads.
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: 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 NewsMore Jev resources like this
Fastino's 287M multilingual classification model in the GLiNER2.5 family: pass any label set at call time and get labels back in a single forward pass, with no prompt template and no generated tokens. It averages 56.7% exact-match on the Fast Decisions suite.
Jev Models & Reimplementations#community#open-models#local
Community
The v3 Jev-Style decision model at 2B in GGUF: F16 3.78 GB and Q8_0 2.01 GB match PyTorch FP32 on 100% and 99.7% of parity rows, 73.6% on JevBench v1.4.1 public items.
Jev Models & Reimplementations#community#open-models#local
Community
Perplexity's 27B decision model rebuilt on Qwen3.8-27B with a joint request format that answers a whole request's questions in one sequence, reporting 68.09 on Decision Index 0.3 and 77.43 on vision.
Jev Models & Reimplementations#community#open-models#multimodal
Community
A specialist full fine-tune of mistralai/Ministral-3-3B-Reasoning-2512 by rony for the LocalLLaMA/typed-decisions benchmark: it reads a case, question and lettered options and softmaxes next-token logits, reporting 0.808 accuracy and 0.0773 KL on all/test.
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
Mithril's open, Jev-shaped typed-decision model on DeBERTa-v3-large: choice up to 255 options, score and noul from one forward pass, 0.854 in-domain with ECE 0.022, trained on public gold labels only.
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.