mohit67890 / imajev

Decisions for real-world cases. 
 Small open models (2B · 4B · 9B) that read the photos, records and text a business already has and answer in the options you set, 
with a probability on each and an explicit can't tell . Your system acts when it is sure and hands the rest to a person. 
 Built by Mohit Garg . I build decision systems like this for operations, support and back-office teams: 
refunds, returns, ticket routing, document and photo checks. Work with me → mohitgarg.ai 
 Live demo · Website · Quickstart · Checked examples · Results · Technical report 
## Independent results
Screenshots of the official leaderboards, captured 28 Sep 2026. Each board is run by its own maintainer, not by us; click through for the live page.
### #1 of 91 on JevBench v1.4.2.2
Scored 27 Sep 2026. imajev-4b 67.37, ahead of Plumb-4B (65.84) and TypeSafe's own Jev 1.13.0 (63.29).
### #1 of 49 on Image JevBench v0.1.3
Released 28 Sep 2026. imajev-4b 76.39, ahead of Jev-Omni (12B, 73.10) and NeoHorse Jev 4B (71.94); imajev-2b is #6. Sealed accuracy 83.3%, second of 44 self-hosted systems. Measured by the maintainer on the fast server (`--fast --merge-lora`, one option order, shipped calibration).
### #1 on JevBench's headline Capability ranking
"Imajev-4B leads the Jev-class systems with 66.3" (Capability: intelligence and calibration, among systems within 2× Jev's cost and latency).
### #3 of 56 models on DecisionBench (eng, v1)
79.65, ahead of GLM-5.3 Flash (320B), Jev 1.13, DeepSeek V4.1 Flash (552B) and GPT-5.6 Luna. The two above are the benchmark team's own Bosun models.
### #3 on DecisionBench Reasoning
80.58, behind GLM-5.3 Flash and GPT-5.6 Luna, ahead of DeepSeek V4.1 Flash (552B) and Jev 1.13.
 Boards move as new systems are added; ranks are quoted with the board version and the date. Archived copies and the raw data are linked in Official leaderboards . 
## Try it live
The live demo runs imajev-4b on a GPU. Pick one of the checked examples
(a listing against its photo, a return, a part on the line, an email against a CRM record, a refund against the policy, a ticket,
a stylist request it declines to guess on), change the record or swap a photo, and watch the answer and the app's action change.
Or upload your own photo and write your own questions.
## One request, every answer typed
Send the evidence and the questions you care about. Each answer comes back as a probability over the answers you allowed, ready
for an `if`. This is the exact script we ran against imajev-4b (`site/showcase/listing.py`) and its output, with numbers rounded to
three places and `usage` shortened; 1.15 s on a Mac Studio (four option orders averaged, calibration file applied).
 [code] 
 Result 
 [code] 
 The same endpoint with no photo: one request routes the ticket (choice), flags urgency (yes/no) and scores frustration (score). Script: site/showcase/ticket.py . 
## What sets it apart
Jev reads text only. General vision models answer in prose, from a hosted API, in seconds. imajev