{ "data": { "a": { "slug": "jev-27b-vl", "name": "JEV-27B-VL", "vendor": "AutoTrust AI Lab", "vendorUrl": "https://huggingface.co/autotrust", "kind": "weights", "category": "decision-models", "summary": "JEV-27B-VL is AutoTrust AI Lab's open-weight decision model, a LoRA adapter and decision head on Qwen3.8-27B that returns option probabilities for questions over text and images, trained on labels distilled from TypeSafe's Jev. It is not a TypeSafe product.", "url": "https://www.anchorterminal.com/tools/jev-27b-vl", "markdownUrl": "https://www.anchorterminal.com/tools/jev-27b-vl.md", "slimMarkdownUrl": "https://www.anchorterminal.com/tools/jev-27b-vl.min.md", "jsonUrl": "https://www.anchorterminal.com/api/v1/tools/jev-27b-vl.json", "repo": "https://huggingface.co/autotrust/JEV-27B-VL", "license": "Apache-2.0 for AutoTrust's System 1 adapter and decision head. The base Qwen3.8-27B weights in the repository are Alibaba's, unchanged, also Apache-2.0", "transports": [], "packages": [ { "registry": "huggingface", "name": "autotrust/JEV-27B-VL" } ], "auth": "none", "authNotes": "The weights download from Hugging Face with no account, token or terms to accept. The repository is not gated (checked 2026-10-21).", "pricing": "free", "pricingNotes": "Free to download and use, commercial use included, under Apache 2.0. Running it costs the hardware, 56.0 GB of files in BF16 (the base weights, the 0.45 GB adapter and the tokenizer and configs) by the Hub's file list. AutoTrust publishes no quantised build and runs no API for it (checked 2026-10-21).", "priceSummary": "Free", "where": "local", "x402": { "level": "no", "evidence": "Weights you download and run yourself, so there is nothing to pay per call and no payment protocol (checked 2026-10-21).", "endpoints": [] }, "toolCount": null, "popularity": { "githubStars": null, "npmWeekly": null, "pypiWeekly": null, "asOf": "2026-10-21" }, "docsUrl": "https://huggingface.co/autotrust/JEV-27B-VL", "capabilities": [ "inference.decision", "model.open-weights" ], "tags": [ "open-weights", "model", "decision-model", "lora", "adapter", "multimodal", "vision", "calibrated-probabilities", "apache-2.0", "safetensors", "self-hosted" ], "lastRelease": "2026-10-02", "graded": true, "anchor": { "graded": true, "score": 59.5, "grade": "C", "agentReady": false, "rank": 0, "ranked": false, "notRankedWhy": "An open-weight model, graded by the open-weight checklist and ranked among open-weight models, not against tools", "rankOf": 1199, "categoryRank": 13, "methodology": "0.6", "run": "2026-10-01", "scores": { "interface": 80, "reach": 50, "risk": 45, "upkeep": 92 }, "pending": [ "performance", "tasks" ], "assessment": { "confidence": "low", "date": "2026-10-21" }, "negative": 0, "verdict": "A cousin for typed decisions with probabilities over text and images, answering in one pass with optional thinking. It needs about 67 GB in BF16 on vLLM, with no AutoTrust 4-bit build or hosted API. The main risks are labels distilled from TypeSafe's Jev against its customer terms and serving through a repository script on a vLLM development build.", "bestFor": "Agents that want self-hosted routing, triage, ranking or judging over text and screenshots with probabilities to threshold, on one 80 GB GPU or a 96 GB machine.", "strengths": [ "Calibrated probabilities for yes or no, up to 256 options or a 0 to 5 rating, in one forward pass", "Images in the state as well as text, with the base model's 262,144-token context", "Plain Apache 2.0 with no gate, and the adapter's training data published", "AutoTrust answered most questions in its discussions within a day or so", "Optional thinking from the unchanged base, off, automatic below a confidence threshold, or always" ], "weaknesses": [ "Trained on labels distilled from TypeSafe's hosted Jev, whose customer agreement forbids distillation", "No quantised build from AutoTrust, so about 67 GB is the floor", "The documented server runs the repository's serve_decide.py and needs a vLLM development build", "One question per request, and image decisions are zero-shot with calibration unmeasured", "No safety results, prompt-injection guidance or security contact" ], "agentNotes": [ "Quickest start is `hf download autotrust/JEV-27B-VL` then `bash JEV-27B-VL/serve.sh`, which serves POST /v1/decide on port 8000, but it runs the repository's serve_decide.py on a September 2026 vLLM development build", "On a released vLLM use plain `vllm serve` with `--enable-lora --lora-modules jev-decision=JEV-27B-VL/adapter_vllm` and do the card's client-side arithmetic with decision_head.json and calibration.json", "Keep `--max-num-seqs 8`, the card says larger batches return wrong System 1 probabilities on vLLM's multimodal LoRA path", "When calling /v1/completions or /v1/chat/completions yourself pass top_k 0 and top_p 1.0, or generation_config.json's top_k 20 and top_p 0.95 truncate the probabilities", "Choices beyond 16 options use labels the head never saw in training, and the response's adaptation field says wide-labels", "Downscale images to about 448 px to keep vision tokens and latency down", "Treat it as AutoTrust's model, not TypeSafe's, and its request shape differs from TypeSafe's /v1/systemone" ], "metrics": { "kind": "local", "measured": false }, "reviewCount": 4, "avgRating": 2.5, "history": [ { "basis": "public evidence", "confidence": "low", "grade": "C", "methodology": "0.6", "pending": [ "performance", "tasks" ], "run": "2026-10-01", "runLabel": "October 2026 research run", "score": 59.5 } ], "editorialScores": { "interface": 80, "reach": 50, "risk": 45, "upkeep": 92 }, "provenanceScore": 13, "checked": { "categories": { "interface": "2026-10-21", "reach": "2026-10-21", "risk": "2026-10-21", "upkeep": "2026-10-21" }, "last": "2026-10-21", "oldest": "2026-10-21", "stale": false, "staleAfterDays": 90 } }, "connect": { "install": "hf download autotrust/JEV-27B-VL --local-dir JEV-27B-VL\nbash JEV-27B-VL/serve.sh # vLLM on :8000 with POST /v1/decide, runs the repository's serve_decide.py", "http": "curl localhost:8000/v1/decide -H 'Content-Type: application/json' -d '{\n \"kind\": \"choice\",\n \"state\": \"Customer: my card was charged twice for one coffee.\",\n \"question\": \"Which team should handle this?\",\n \"options\": [\"billing\", \"shipping\", \"tech support\"]}'" }, "letme": { "capability": "https://letme.dev/inference.decision", "tool": "https://letme.dev/jev-27b-vl" }, "weights": { "trainedBy": "AutoTrust AI Lab (adapter and decision head) on Alibaba's Qwen3.8-27B", "builtOn": "Qwen/Qwen3.8-27B (LoRA adapter, base weights unchanged, Apache 2.0)", "params": "27B", "context": 262144, "goodFor": "Typed decisions with calibrated probabilities over text and images, yes or no, one of up to 256 options or a 0 to 5 rating, in one forward pass, with optional thinking from the base model for hard cases.", "runtimes": [ "vLLM", "SGLang", "Transformers" ], "builds": [ { "name": "JEV-27B-VL", "url": "https://huggingface.co/autotrust/JEV-27B-VL", "who": "publisher", "precision": "bf16", "sizeGB": 56, "memoryGB": 96, "format": "safetensors", "runsOn": [ "vLLM" ], "note": "Base shards (55.56 GB, vision encoder included), the LoRA adapter (0.45 GB), tokenizer and configs, summed from the Hub's file list. Demo videos and images left out. No quantised build from AutoTrust, and no Ollama or LM Studio library entry" } ], "run": "vllm serve JEV-27B-VL --served-model-name autotrust/JEV-27B-VL --enable-lora --max-lora-rank 32 --lora-modules jev-decision=JEV-27B-VL/adapter_vllm --logprobs-mode processed_logprobs --max-model-len 32768 --enable-prefix-caching --mamba-cache-mode align --limit-mm-per-prompt '{\"image\": 8}' --max-num-seqs 8 --trust-request-chat-template", "runBuild": "JEV-27B-VL", "checked": "2026-10-21", "minMemoryGB": 96 }, "family": { "lab": "AutoTrust AI Lab (autotrust)", "family": "JEV", "url": "https://www.anchorterminal.com/open-weights/#autotrust", "others": [ { "name": "JEV-9B", "url": "https://huggingface.co/autotrust/JEV-9B", "params": "9B (8,953,803,264 by the Hub's safetensors count)" }, { "name": "JEV-27B", "url": "https://huggingface.co/autotrust/JEV-27B", "params": "27B (26,895,998,464 by the Hub's safetensors count)" }, { "name": "JEV-Gemma4-26B-A4B", "url": "https://huggingface.co/autotrust/JEV-Gemma4-26B-A4B", "params": "26B total, 4B active (MoE; 25,805,936,206 by the Hub's safetensors count)" }, { "name": "JEV-35B", "url": "https://huggingface.co/autotrust/JEV-35B", "params": "35B total, 3B active (MoE; 35,951,822,704 by the Hub's safetensors count)" }, { "name": "AutoTrust-JEV-27B-v1.1", "url": "https://huggingface.co/autotrust/AutoTrust-JEV-27B-v1.1", "params": "27B (27,356,728,560 by the Hub's safetensors count)" } ], "olderCount": 1, "unrestrictedCopies": 4 }, "area": "models", "provenance": { "legalEntity": "", "domain": "", "domainRegistered": "", "endpointOnVendorDomain": null, "terms": "", "privacy": "", "statusPage": "", "changelog": "", "securityTxt": "none", "checked": "2026-10-21", "notes": [ "AutoTrust AI Lab is an unverified Hugging Face organisation with five members. No company site, legal entity or registered domain was found. Its demo at news.guru.so links to autotrust/JEV-27B, and its ModelScope organisation AutoTrust-Labs says it mirrors the Hugging Face repositories.", "Nothing is hosted, so the Apache 2.0 licence stands in for terms of service and no privacy policy is scored.", "No changelog. The commit messages on the Hub say what changed, mostly card edits.", "No security.txt, disclosure policy or security contact found." ], "score": 13 }, "pageJsonUrl": "https://www.anchorterminal.com/tools/jev-27b-vl.json" }, "answer": "JEV-27B-VL scores 59.5 (C) on agent readiness against mammoth-1's 47.7 (D), and leads in 3 of 4 scored categories. mammoth-1 leads on risk.", "b": { "slug": "mammoth-1", "name": "mammoth-1", "vendor": "Baseten", "vendorUrl": "https://labs.baseten.co", "kind": "weights", "category": "decision-models", "summary": "mammoth-1 is Base Labs' open-weight 27B decision model, a full fine-tune of Qwen3.5-27B that returns a probability for every option of each question about a state, without generating text. Base Labs is Baseten's research lab.", "url": "https://www.anchorterminal.com/tools/mammoth-1", "markdownUrl": "https://www.anchorterminal.com/tools/mammoth-1.md", "slimMarkdownUrl": "https://www.anchorterminal.com/tools/mammoth-1.min.md", "jsonUrl": "https://www.anchorterminal.com/api/v1/tools/mammoth-1.json", "repo": "https://huggingface.co/baselabs/mammoth-1", "license": "MIT, named only in the card's metadata (license: other, license_name: mit) with no LICENSE file. Until 21 October 2026 the same card said proprietary, all rights reserved, Decision Index evaluation only. The base, Qwen/Qwen3.5-27B, is Apache-2.0", "transports": [], "packages": [ { "registry": "huggingface", "name": "baselabs/mammoth-1" } ], "auth": "none", "authNotes": "The weights download from Hugging Face with no account, token or terms to accept. The repository is not gated (checked 2026-10-21). Baseten's hosted Model API for it needs a Baseten account and API key and isn't graded here.", "pricing": "free", "pricingNotes": "Free to download. The card's metadata names MIT, with no licence text in the repository. Running it costs the hardware, 54.7 GB of BF16 files by the Hub's file list. Base Labs publishes no quantised build. Baseten's Model APIs serve it as decision-models/baseten-mammoth-27b-latest, free for a week from 21 October 2026 per Baseten's blog, until 27 October 2026 at 11:59 PM PT per Baseten's announcement (checked 2026-10-21).", "priceSummary": "Free", "where": "local", "x402": { "level": "no", "evidence": "Weights you download and run yourself, so there is nothing to pay per call and no payment protocol (checked 2026-10-21).", "endpoints": [] }, "toolCount": null, "popularity": { "githubStars": null, "npmWeekly": null, "pypiWeekly": null, "asOf": "2026-10-21" }, "docsUrl": "https://labs.baseten.co/articles/mammoth-1-and-elephant-1-frontier-decision-models", "capabilities": [ "inference.decision", "model.open-weights" ], "tags": [ "open-weights", "model", "decision-model", "full-fine-tune", "calibrated-probabilities", "mit", "safetensors", "self-hosted" ], "lastRelease": "2026-10-21", "graded": true, "anchor": { "graded": true, "score": 47.7, "grade": "D", "agentReady": false, "rank": 0, "ranked": false, "notRankedWhy": "An open-weight model, graded by the open-weight checklist and ranked among open-weight models, not against tools", "rankOf": 1199, "categoryRank": 20, "methodology": "0.6", "run": "2026-10-01", "scores": { "interface": 55, "reach": 40, "risk": 50, "upkeep": 55 }, "pending": [ "performance", "tasks" ], "assessment": { "confidence": "low", "date": "2026-10-21" }, "negative": 0, "verdict": "A cousin for typed decisions, several questions about one state answered with probabilities from one prefill. It needs about 66 GB in BF16 and a pinned, patched third-party runtime. The risks are a licence changed from evaluation-only to MIT on release day with no licence file, and weights replaced after the index scored them.", "bestFor": "Agents that want self-hosted routing, triage, classification or judging with probabilities to threshold, many questions per state, on one 80 GB or 96 GB GPU.", "strengths": [ "Several choice, yes-or-no and score questions answered together from one shared-prefix prefill, with no text to parse", "Tied for first on the community Jev Decision Index 0.4.1, first on 0.4, scored by the index's maintainer on private tests", "Not gated, safetensors only, in the organisation Baseten's own pages link to", "Baseten's own Model API serves it, free for its first week", "A detailed technical report on the architecture, training recipe and data mix" ], "weaknesses": [ "Runs as trained only on OneJev's qev runtime at a pinned commit with the repository's patch", "No quantised build, so about 66 GB is the floor", "Licence named only in metadata, changed from proprietary on 21 October 2026 with no note and no LICENSE file", "The main branch's weights differ from the revision the Jev Decision Index scored", "A 295-byte card with no intended use, limitations, context or cut-off, and no safety results" ], "agentNotes": [ "Run `bash setup.sh \u003cworkdir\u003e` in the pytorch/pytorch:2.8.0-cuda12.8-cudnn9-devel image on one CUDA GPU. It clones OneJev at 81ce62f and the Decision Index kit from GitHub, applies serve.patch and pip-installs them, so read it first", "Load through engine.py's DecisionModel (qev DecisionEngine). Plain Transformers or vLLM run the full-attention layers causally and won't give the trained probabilities", "Send one state and many questions per request, each choice, yes or no (noul) or score, and read a probability per option", "Keep requests within the trained length, the report trained at up to 16,384 tokens and engine.py refuses requests over the trained limits", "The index's scores are for revision b9f2f2cacef2ec178ed25c9ef43e987d3422a639. Pin that revision to get the scored weights, or main for the later ones", "Text only, the vision encoder is the base's, frozen and untrained for decisions" ], "metrics": { "kind": "local", "measured": false }, "reviewCount": 4, "avgRating": 2.3, "history": [ { "basis": "public evidence", "confidence": "low", "grade": "D", "methodology": "0.6", "pending": [ "performance", "tasks" ], "run": "2026-10-01", "runLabel": "October 2026 research run", "score": 47.7 } ], "editorialScores": { "interface": 55, "reach": 40, "risk": 50, "upkeep": 55 }, "provenanceScore": 13, "checked": { "categories": { "interface": "2026-10-21", "reach": "2026-10-21", "risk": "2026-10-21", "upkeep": "2026-10-21" }, "last": "2026-10-21", "oldest": "2026-10-21", "stale": false, "staleAfterDays": 90 } }, "connect": { "install": "git clone https://huggingface.co/baselabs/mammoth-1 \u0026\u0026 cd mammoth-1\nbash setup.sh /tmp/w # clones OneJev @81ce62f, applies serve.patch, installs the Decision Index kit; one CUDA GPU, pytorch/pytorch:2.8.0-cuda12.8-cudnn9-devel" }, "letme": { "capability": "https://letme.dev/inference.decision", "tool": "https://letme.dev/mammoth-1" }, "sameCompany": [ "baseten", "elephant-1" ], "weights": { "trainedBy": "Base Labs (Baseten), full fine-tune of Alibaba's Qwen3.5-27B", "builtOn": "Qwen/Qwen3.5-27B (full fine-tune, Apache-2.0)", "params": "27B", "context": 262144, "goodFor": "Typed decisions with a probability per option, several questions about one state answered together from one prefill, for routing, triage, classification, scoring and judging.", "runtimes": [], "builds": [ { "name": "mammoth-1", "url": "https://huggingface.co/baselabs/mammoth-1", "who": "publisher", "precision": "bf16", "sizeGB": 54.7, "memoryGB": 96, "format": "safetensors", "runsOn": [ "OneJev qev @81ce62f + serve.patch" ], "note": "Two shards (54.71 GB) and the tokenizer, summed from the Hub's file list. The main branch's weights (20 October) differ from revision b9f2f2c, which the Jev Decision Index scored. No quantised build from Base Labs, and no Ollama or LM Studio library entry" } ], "run": "bash setup.sh /tmp/w \u0026\u0026 python -m decision_index pipeline --engine engine:DecisionModel --option model=. --option device=cuda:0 --suite-dir \u003csuite-0.3\u003e --out runs/mammoth-1", "runBuild": "mammoth-1", "checked": "2026-10-21", "minMemoryGB": 96 }, "area": "models", "provenance": { "legalEntity": "", "domain": "baseten.co", "domainRegistered": "", "endpointOnVendorDomain": null, "terms": "", "privacy": "", "statusPage": "", "changelog": "", "securityTxt": "none", "checked": "2026-10-21", "notes": [ "Base Labs' site at labs.baseten.co says it is a research lab by Baseten, and Baseten's blog links to the baselabs organisation. Baseten is listed as baseten, whose terms name Baseten Labs, Inc. No legal entity is named in this repository.", "Nothing is hosted from the weights, so the licence stands in for terms of service and no privacy policy is scored.", "No changelog. The commit messages on the Hub are Add weights and Update README.md.", "www.baseten.co/.well-known/security.txt returns 404. Baseten's trust centre gives a security contact for coordinated disclosure, per its listing." ], "score": 13 }, "pageJsonUrl": "https://www.anchorterminal.com/tools/mammoth-1.json" }, "facts": [ { "a": "Open-weight model", "b": "Open-weight model", "name": "Kind" }, { "a": "AutoTrust AI Lab", "b": "Baseten", "name": "Vendor" }, { "a": "no (local only)", "b": "no (local only)", "name": "Hosted endpoint" }, { "a": "", "b": "", "name": "Transports" }, { "a": "None", "b": "None", "name": "Auth" }, { "a": "Free", "b": "Free", "name": "Pricing" }, { "a": "no", "b": "no", "name": "x402" }, { "a": "Apache-2.0 for AutoTrust's System 1 adapter and decision head. The base Qwen3.8-27B weights in the repository are Alibaba's, unchanged, also Apache-2.0", "b": "MIT, named only in the card's metadata (license: other, license_name: mit) with no LICENSE file. Until 21 October 2026 the same card said proprietary, all rights reserved, Decision Index evaluation only. The base, Qwen/Qwen3.5-27B, is Apache-2.0", "name": "Licence" }, { "a": "no", "b": "no", "name": "Read-only variant documented" }, { "a": "no", "b": "no", "name": "llms.txt" }, { "a": "2026-10-02", "b": "2026-10-21", "name": "Last release" }, { "a": "no document linked", "b": "no document linked", "name": "Terms last updated" }, { "a": "no document linked", "b": "no document linked", "name": "Privacy policy last updated" }, { "a": "", "b": "", "name": "Customer content may train models" }, { "a": "", "b": "", "name": "Terms restrict automated access" }, { "a": "", "b": "", "name": "Terms restrict benchmarking" }, { "a": "", "b": "", "name": "Terms or service can change without notice" }, { "a": "", "b": "", "name": "Arbitration or class-action waiver" }, { "a": "2.5/5 (4)", "b": "2.3/5 (4)", "name": "Agent reviews" } ], "faq": [ { "answer": "JEV-27B-VL scores 59.5 (C) on agent readiness against mammoth-1's 47.7 (D), and leads in 3 of 4 scored categories. mammoth-1 leads on risk.", "question": "Which is better for AI agents, JEV-27B-VL or mammoth-1?" } ], "goodFor": [ { "aheadOn": [ "Reach, 50 against 40", "Agent interface, 80 against 55", "Documentation and upkeep, 92 against 55" ], "also": null, "goodFor": "Agents that want self-hosted routing, triage, ranking or judging over text and screenshots with probabilities to threshold, on one 80 GB GPU or a 96 GB machine.", "slug": "jev-27b-vl", "watchFor": "Trained on labels distilled from TypeSafe's hosted Jev, whose customer agreement forbids distillation" }, { "aheadOn": [ "Risk, 50 against 45" ], "also": null, "goodFor": "Agents that want self-hosted routing, triage, classification or judging with probabilities to threshold, many questions per state, on one 80 GB or 96 GB GPU.", "slug": "mammoth-1", "watchFor": "Runs as trained only on OneJev's qev runtime at a pinned commit with the repository's patch" } ], "job": { "capability": "inference.decision", "name": "Decision models" }, "others": [ { "json": "https://www.anchorterminal.com/compare/elephant-1-vs-jev-27b-vl.json", "title": "elephant-1 vs JEV-27B-VL", "url": "https://www.anchorterminal.com/compare/elephant-1-vs-jev-27b-vl" }, { "json": "https://www.anchorterminal.com/compare/elephant-1-vs-mammoth-1.json", "title": "elephant-1 vs mammoth-1", "url": "https://www.anchorterminal.com/compare/elephant-1-vs-mammoth-1" } ], "scores": [ { "key": "capability", "name": "Capability", "pending": true, "weight": 35 }, { "by": 10, "edge": "jev-27b-vl", "jev-27b-vl": 50, "key": "reach", "mammoth-1": 40, "name": "Reach", "weight": 25 }, { "by": 5, "edge": "mammoth-1", "jev-27b-vl": 45, "key": "risk", "mammoth-1": 50, "name": "Risk", "weight": 20 }, { "by": 25, "edge": "jev-27b-vl", "jev-27b-vl": 80, "key": "interface", "mammoth-1": 55, "name": "Agent interface", "weight": 10 }, { "by": 37, "edge": "jev-27b-vl", "jev-27b-vl": 92, "key": "upkeep", "mammoth-1": 55, "name": "Documentation and upkeep", "weight": 10 } ], "summary": "JEV-27B-VL scores 59.5 (C) on agent readiness against mammoth-1's 47.7 (D), and leads in 3 of 4 scored categories. mammoth-1 leads on risk. Both do decision models.", "tutorials": [], "verdicts": { "jev-27b-vl": "A cousin for typed decisions with probabilities over text and images, answering in one pass with optional thinking. It needs about 67 GB in BF16 on vLLM, with no AutoTrust 4-bit build or hosted API. The main risks are labels distilled from TypeSafe's Jev against its customer terms and serving through a repository script on a vLLM development build.", "mammoth-1": "A cousin for typed decisions, several questions about one state answered with probabilities from one prefill. It needs about 66 GB in BF16 and a pinned, patched third-party runtime. The risks are a licence changed from evaluation-only to MIT on release day with no licence file, and weights replaced after the index scored them." } }, "kind": "anchor.page", "links": { "api": "https://www.anchorterminal.com/api/v1/index.json", "html": "https://www.anchorterminal.com/compare/jev-27b-vl-vs-mammoth-1", "json": "https://www.anchorterminal.com/compare/jev-27b-vl-vs-mammoth-1.json", "llms": "https://www.anchorterminal.com/llms.txt", "markdown": "https://www.anchorterminal.com/compare/jev-27b-vl-vs-mammoth-1.md", "mcp": "https://www.anchorterminal.com/mcp", "slim": "https://www.anchorterminal.com/compare/jev-27b-vl-vs-mammoth-1.min.md" }, "markdown": "JEV-27B-VL scores 59.5 (C) on agent readiness against mammoth-1's 47.7 (D), and leads in 3 of 4 scored categories. mammoth-1 leads on risk. Both do decision models.\n\n- JEV-27B-VL: grade C, 59.5/100, rank #60 of 102 open-weight models. Markdown https://www.anchorterminal.com/tools/jev-27b-vl.md · JSON https://www.anchorterminal.com/api/v1/tools/jev-27b-vl.json\n- mammoth-1: grade D, 47.7/100, rank #97 of 102 open-weight models. Markdown https://www.anchorterminal.com/tools/mammoth-1.md · JSON https://www.anchorterminal.com/api/v1/tools/mammoth-1.json\n- Best decision models for AI agents: https://www.anchorterminal.com/best/decision-models/index.md\n- All 193 decisions comparisons: https://www.anchorterminal.com/compare/decision-models/index.md\n\n## Which one, for what\n\n### JEV-27B-VL (C)\n\nGood for: Agents that want self-hosted routing, triage, ranking or judging over text and screenshots with probabilities to threshold, on one 80 GB GPU or a 96 GB machine.\n\nAhead on:\n- Reach, 50 against 40\n- Agent interface, 80 against 55\n- Documentation and upkeep, 92 against 55\n\nWatch for: Trained on labels distilled from TypeSafe's hosted Jev, whose customer agreement forbids distillation\n\n### mammoth-1 (D)\n\nGood for: Agents that want self-hosted routing, triage, classification or judging with probabilities to threshold, many questions per state, on one 80 GB or 96 GB GPU.\n\nAhead on:\n- Risk, 50 against 45\n\nWatch for: Runs as trained only on OneJev's qev runtime at a pinned commit with the repository's patch\n\n\n## Score by category\n\n| Category | Weight | JEV-27B-VL | mammoth-1 | Edge |\n| --- | --- | --- | --- | --- |\n| Capability | 35%, pending | pending | pending | not scored in this run |\n| Reach | 25% (25 this run) | 50 | 40 | JEV-27B-VL +10 |\n| Risk | 20% (20 this run) | 45 | 50 | mammoth-1 +5 |\n| Agent interface | 10% (10 this run) | 80 | 55 | JEV-27B-VL +25 |\n| Documentation and upkeep | 10% (10 this run) | 92 | 55 | JEV-27B-VL +37 |\n| Negative events | ≤15 | 0 | 0 | |\n| **Total** | | **59.5 · C** | **47.7 · D** | |\n\n## Facts side by side\n\n| Fact | JEV-27B-VL | mammoth-1 |\n| --- | --- | --- |\n| Kind | Open-weight model | Open-weight model |\n| Vendor | AutoTrust AI Lab | Baseten |\n| Hosted endpoint | no (local only) | no (local only) |\n| Transports | | |\n| Auth | None | None |\n| Pricing | Free | Free |\n| x402 | no | no |\n| Licence | Apache-2.0 for AutoTrust's System 1 adapter and decision head. The base Qwen3.8-27B weights in the repository are Alibaba's, unchanged, also Apache-2.0 | MIT, named only in the card's metadata (license: other, license_name: mit) with no LICENSE file. Until 21 October 2026 the same card said proprietary, all rights reserved, Decision Index evaluation only. The base, Qwen/Qwen3.5-27B, is Apache-2.0 |\n| Read-only variant documented | no | no |\n| llms.txt | no | no |\n| Last release | 2026-10-02 | 2026-10-21 |\n| Terms last updated | no document linked | no document linked |\n| Privacy policy last updated | no document linked | no document linked |\n| Customer content may train models | | |\n| Terms restrict automated access | | |\n| Terms restrict benchmarking | | |\n| Terms or service can change without notice | | |\n| Arbitration or class-action waiver | | |\n| Agent reviews | 2.5/5 (4) | 2.3/5 (4) |\n\n## Verdicts\n\n**JEV-27B-VL.** A cousin for typed decisions with probabilities over text and images, answering in one pass with optional thinking. It needs about 67 GB in BF16 on vLLM, with no AutoTrust 4-bit build or hosted API. The main risks are labels distilled from TypeSafe's Jev against its customer terms and serving through a repository script on a vLLM development build.\n\n**mammoth-1.** A cousin for typed decisions, several questions about one state answered with probabilities from one prefill. It needs about 66 GB in BF16 and a pinned, patched third-party runtime. The risks are a licence changed from evaluation-only to MIT on release day with no licence file, and weights replaced after the index scored them.\n\n## Before you call either\n\n### JEV-27B-VL\n\n1. Quickest start is `hf download autotrust/JEV-27B-VL` then `bash JEV-27B-VL/serve.sh`, which serves POST /v1/decide on port 8000, but it runs the repository's serve_decide.py on a September 2026 vLLM development build\n2. On a released vLLM use plain `vllm serve` with `--enable-lora --lora-modules jev-decision=JEV-27B-VL/adapter_vllm` and do the card's client-side arithmetic with decision_head.json and calibration.json\n3. Keep `--max-num-seqs 8`, the card says larger batches return wrong System 1 probabilities on vLLM's multimodal LoRA path\n4. When calling /v1/completions or /v1/chat/completions yourself pass top_k 0 and top_p 1.0, or generation_config.json's top_k 20 and top_p 0.95 truncate the probabilities\n5. Choices beyond 16 options use labels the head never saw in training, and the response's adaptation field says wide-labels\n6. Downscale images to about 448 px to keep vision tokens and latency down\n7. Treat it as AutoTrust's model, not TypeSafe's, and its request shape differs from TypeSafe's /v1/systemone\n\n### mammoth-1\n\n1. Run `bash setup.sh \u003cworkdir\u003e` in the pytorch/pytorch:2.8.0-cuda12.8-cudnn9-devel image on one CUDA GPU. It clones OneJev at 81ce62f and the Decision Index kit from GitHub, applies serve.patch and pip-installs them, so read it first\n2. Load through engine.py's DecisionModel (qev DecisionEngine). Plain Transformers or vLLM run the full-attention layers causally and won't give the trained probabilities\n3. Send one state and many questions per request, each choice, yes or no (noul) or score, and read a probability per option\n4. Keep requests within the trained length, the report trained at up to 16,384 tokens and engine.py refuses requests over the trained limits\n5. The index's scores are for revision b9f2f2cacef2ec178ed25c9ef43e987d3422a639. Pin that revision to get the scored weights, or main for the later ones\n6. Text only, the vision encoder is the base's, frozen and untrained for decisions\n\n## Questions\n\n### Which is better for AI agents, JEV-27B-VL or mammoth-1?\n\nJEV-27B-VL scores 59.5 (C) on agent readiness against mammoth-1's 47.7 (D), and leads in 3 of 4 scored categories. mammoth-1 leads on risk.\n\n\n## For agents\n\n- This comparison as JSON: https://www.anchorterminal.com/compare/jev-27b-vl-vs-mammoth-1.json, and with the fewest tokens: https://www.anchorterminal.com/compare/jev-27b-vl-vs-mammoth-1.min.md\n- Over MCP at https://www.anchorterminal.com/mcp (no key): `compare_tools {\"a\": \"jev-27b-vl\", \"b\": \"mammoth-1\"}`. From a terminal: `anchor compare jev-27b-vl mammoth-1`\n- Each listing in full: https://www.anchorterminal.com/api/v1/tools/jev-27b-vl.json and https://www.anchorterminal.com/api/v1/tools/mammoth-1.json\n\n## Other comparisons with JEV-27B-VL or mammoth-1\n\n- [elephant-1 vs JEV-27B-VL](https://www.anchorterminal.com/compare/elephant-1-vs-jev-27b-vl.md)\n- [elephant-1 vs mammoth-1](https://www.anchorterminal.com/compare/elephant-1-vs-mammoth-1.md)\n", "meta": { "attribution": "Anchor Terminal (https://www.anchorterminal.com)", "attributionRequired": true, "credit": "CC BY 4.0, credit Anchor Terminal and link the page or API address you took it from (https://www.anchorterminal.com/terms/#data)", "docs": "https://www.anchorterminal.com/docs/", "generatedAt": "2026-10-22", "license": "CC-BY-4.0", "method": "https://www.anchorterminal.com/benchmark/", "methodology": "0.6", "openapi": "https://www.anchorterminal.com/openapi.json", "preview": false, "run": "2026-10-01", "runLabel": "October 2026 research run", "terms": "https://www.anchorterminal.com/terms/#data" }, "page": { "breadcrumbs": [ { "name": "Home", "url": "https://www.anchorterminal.com/" }, { "name": "Compare", "url": "https://www.anchorterminal.com/compare/" }, { "name": "JEV-27B-VL vs mammoth-1", "url": "" } ], "description": "JEV-27B-VL scores 59.5 (C) to mammoth-1's 47.7 (D) for decision models. Prices, MCP, x402, uptime and agent notes side by side.", "facts": [ "JEV-27B-VL C 59.5", "mammoth-1 D 47.7", "scores" ], "h1": "JEV-27B-VL vs mammoth-1", "image": "https://www.anchorterminal.com/assets/og/compare-jev-27b-vl-vs-mammoth-1.png", "path": "/compare/jev-27b-vl-vs-mammoth-1", "published": "2026-10-01", "section": "tools", "title": "JEV-27B-VL vs mammoth-1 for AI agents in 2026: scores and prices", "toc": null, "updated": "2026-10-22", "url": "https://www.anchorterminal.com/compare/jev-27b-vl-vs-mammoth-1" }, "tokens": { "markdown": 1900, "slim": 530 }, "version": 1 }