OpenJev. System One model ready for any workload: images, videos, texts

They are blind.

13 of 14 against 6. A model that cannot see the parcel still tells your claims pipeline, with a probability attached, that the parcel is fine.

The other models take text only and ignore the image field. Photos are from Wikimedia Commons, credited at the end.

Video, too.

Send a clip in video_data and ask the same kinds of question. The clip is cut into six evenly spaced frames, every frame answers, and an option is as likely as its best frame makes it. You get the decision and the answer of every frame with its timestamp, so you also know when.

Six real clips from Wikimedia Commons, 7 to 25 seconds, eight questions, two of them asking for something that is not there. A request takes 2 to 12 seconds on the 4B; the first clip at a new resolution took 33. The text-only models have nowhere to put a video at all.

Jev's benchmark. Jev's tasks. Jev's wrong answers.

There are exactly 24 hours until departure. Free cancellation is allowed at or before that 24-hour cutoff.

JevBench is public, and so is every answer Jev gave on it. We took public tasks where OpenJev is right and Jev is wrong, and sent them again, live, to Jev and to everyone else.

A selection from the public JevBench tasks, rerun live.

Faster than all of them.

Twenty short text requests to each model, one at a time, from one laptop, a new connection every time.

OpenJev answers in 228 ms on three V100s, a GPU that shipped in 2017. Jev takes 324. Solar Decide takes 715 and, one request in ten, more than six seconds.

Ours through our public endpoint, the others through OpenRouter.

What you get, and what they charge for.

Same score, a quarter of the price

Everyday decisions are routing, policy checks and triage: the easy and standard tiers of JevBench, 120 public tasks. OpenJev gets 119 of them. So does Jev, at four times the price.

Ranked systems on JevBench v1.2, public easy and standard tasks. Prices of other systems are the benchmark's cost column; ours is our list price after the free 1,000 requests a day.

How it works.

Nothing is generated. Every option of every question becomes a hypothesis, the state is the premise, and a Qwen3.5 cross-encoder reads the pair once and returns three logits: contradiction, entailment, neutral. The probability of an option is its entailment score, normalised over the options of that question. For an image the premise is the pixels.

There is no sampling and no output to parse, so there is nothing to retry. Asked the same thing twice, it gave the same label all 40 times out of 40. The order of your options cannot change the answer: on all 231 JevBench tasks, reversed and shuffled, not one label moved.

The training recipe

Now the part most model cards bury. The v5 panel contains the train and the test splits of MMLU, ARC, GSM8K, HellaSwag, WinoGrande, GPQA, CLINC-150, Banking77 and ESCI. We put them there on purpose, to teach the format of multiple choice, and that makes any score on those sets worthless. We do not report them. No JevBench item was ever in a training mixture: the data builder refuses them by name. The same holds for LLM-AggreFact, HaluBench, the RAGTruth test set, IFEval and LLMBar.

Ablations

Fun demos.

A model that turns a state and a list of options into a decision is a game controller. Nobody trained it to play anything.

Doom, from the pixels

The frame is the premise, statements about where the nearest monster stands are the options, and the most entailed one picks the move: turn left, turn right or fire. 13 kills per episode, 18 in the best one, zero-shot. Random play gets 1.

Minecraft, to an iron pickaxe

A backward-chaining scaffold asks yes-or-no questions about the inventory and the world, and the model only answers them. From an empty inventory to an iron pickaxe: 11 milestones in about 22 decisions.

Doom is played by the v5 model we serve. The Minecraft run was recorded with its v2 checkpoint. The code for both is in the model repository.

Know the limits.

Do not make it the only thing standing between an agent and rm -rf. Put it next to deterministic checks, not in place of them.

Start.

Create a key under settings → API keys. Code written for Jev on OpenRouter needs three changes: the base URL, the key, and the model name.

Three decisions in one request

curl https://www.rawmodels.ai/api/alpha/decisions \
  -H "Authorization: Bearer $RAWMODELS_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "decide",
    "state": {
      "customer_tier": "enterprise",
      "ticket": "My checkout page shows a blank screen after I click Pay. I have tried two browsers."
    },
    "questions": {
      "is_bug": {
        "type": "noul",
        "instructions": "Is the customer reporting a software defect?",
        "criteria": {
          "true": "The customer describes broken or unexpected product behavior.",
          "false": "The customer is asking a question or requesting a feature."
        }
      },
      "team": {
        "type": "choice",
        "instructions": "Which team should own this ticket?",
        "criteria": {
          "payments": "Checkout, billing, or payment processing issues.",
          "frontend": "Rendering, layout, or browser compatibility issues.",
          "account": "Login, permissions, or profile issues."
        }
      },
      "urgency": {
        "type": "score",
        "instructions": "How urgent is this ticket?",
        "criteria": ["Can wait for the next release", "Should be fixed this week", "Blocking revenue right now"]
      }
    }
  }'

{
  "model": "decide",
  "provider": "rawmodels",
  "answers": {
    "is_bug": { "type": "noul", "noul": 0.9689 },
    "team": {
      "type": "choice", "choice": "payments", "confidence": 0.4445,
      "probabilities": { "payments": 0.7039, "frontend": 0.2958, "account": 0.0004 }
    },
    "urgency": {
      "type": "score", "score": 1.9437, "confidence": 0.8004,
      "probabilities": { "0": 0.0017, "1": 0.0529, "2": 0.9454 },
      "legend": {
        "0": "Can wait for the next release",
        "1": "Should be fixed this week",
        "2": "Blocking revenue right now"
      }
    }
  },
  "usage": { "input_tokens": 487, "output_tokens": 0, "cost": 0 }
}

noul returns the probability of yes, choice returns the winning option and the whole distribution, score returns a position on an ordered scale.

A photo

import base64, os, requests

image = base64.b64encode(open("parcel.jpg", "rb").read()).decode()
r = requests.post(
    "https://www.rawmodels.ai/api/alpha/decisions",
    headers={"Authorization": f"Bearer {os.environ['RAWMODELS_API_KEY']}"},
    json={
        "model": "decide",
        "state": "An image: <<IMG>>",
        "image_data": image,
        "questions": {
            "damaged": {
                "type": "noul",
                "instructions": "Is the parcel visibly damaged?",
                "criteria": {
                    "true": "The packaging is torn, crushed or dented.",
                    "false": "The packaging is intact.",
                },
            },
        },
    },
)
print(r.json()["answers"])

A video

video = base64.b64encode(open("dashcam.mp4", "rb").read()).decode()
r = requests.post(
    "https://www.rawmodels.ai/api/alpha/decisions",
    headers={"Authorization": f"Bearer {os.environ['RAWMODELS_API_KEY']}"},
    json={
        "model": "decide",
        "state": "A video frame: <<IMG>>",
        "video_data": video,
        "questions": {
            "hit": {
                "type": "noul",
                "instructions": "Is something hitting the windshield?",
                "criteria": {
                    "true": "Rain, ice or debris is striking the glass.",
                    "false": "The windshield is clear and dry.",
                },
            },
        },
    },
)
print(r.json()["answers"])   # one decision for the clip
print(r.json()["frames"])    # and the answer of every frame, with its second

The full reference is in the Decisions API guide.

Photo and video credits

• Damaged fragile parcel delivered to doorstep, Meanwell Packaging, CC BY 2.0. Resized or cut into frames.
• Black car with flat tire parked on gravel, Shixart1985, CC BY 2.0. Resized or cut into frames.
• Parking ticket, ThorRune, Public domain. Resized or cut into frames.
• Wet floor sign CANEX, Dvermeirre, CC0. Resized or cut into frames.
• Construction workers not wearing fall protection equipment, NIOSH, Public domain. Resized or cut into frames.
• WIC, Seattle, Washington, USDA, Public domain. Resized or cut into frames.
• Shattered light fixture 3, W.carter, CC0. Resized or cut into frames.
• Vertical Stop sign in Moscow, Vlsr1, CC0. Resized or cut into frames.
• Fire extinguisher in wall mount, Mibrahim74, CC BY 3.0. Resized or cut into frames.
• A pothole in Dilova Street in Kyiv, Роман Рябенко, CC0. Resized or cut into frames.
• Montreal Double Red Traffic Light, Emma0mb, CC BY 4.0. Resized or cut into frames.
• Thermometer (Fever), Alabama Extension, CC0. Resized or cut into frames.
• Blue candle burning (video), Jahobr, CC0. Resized or cut into frames.
• Video of tabletop fireplace burning (video), Pittigrilli, CC BY 4.0. Resized or cut into frames.
• Cat jumping backwards (video), Mary Qin, CC BY 3.0. Resized or cut into frames.
• Gigaset Cordless Telephone Production VII (video), Kathinka Engels and others, CC BY 3.0. Resized or cut into frames.
• Freezing rain on the front windows (video), Fumikas Sagisavas, CC BY 4.0. Resized or cut into frames.
• The Velenje Coal Mine, running of the chain conveyor (video), Sounds of Changes, CC BY 3.0. Resized or cut into frames.