VisionHouse model
d3-lite
by vLLM Semantic Router
The lite chassis. Put it first and the soup keeps its answer layer.
Net 1.7 GB · Apache-2.0
01MergeRuns on your machine
Make a soup
Pick up to five decision models that share a base, give each a portion, and choose how to stir. One command cooks the merge on your own machine.
- d3-liteFirst: its files are kept50%
- Intern DecisionVision50%
Ratio 1:1
Stirred evenly: a weighted average of every weight.
Your order
- d3-lite50%
- Intern Decision50%
Linear, Lite · 0.8Babout 3.4 GB to download
npx @zils/cli merge vllm-sr/d3-lite internlm/Intern-Decision-0.8B --out ./my-mergeCooks on your machine with Node.js 20+ and Python 3.11+.
The market
VisionHouse model
Intern Decision
by InternLM
Structured decisions over text and images.
Net 1.7 GB · Apache-2.0
GeneralMarket
jiwo
by eljiwo
A text-only decision fine-tune with a d3-style answer layer.
Net 1.5 GB · Apache-2.0
IT and securityMarket
Cribl Decision 2
by Cribl
Decisions for IT operations and security data.
Net 1.5 GB · Apache-2.0
On deviceMarket
Matilda Edge
by Maincode
Maincode’s smallest decision model, for phones and high throughput.
Net 1.7 GB · Apache-2.0
ScreensMarket
OneJev 1.6
by OmniJev
Decisions over screenshots, photos and video.
Net 1.7 GB · Apache-2.0
GeneralMarket
decider
by Mapika
Reads a state and answers typed questions; text only.
Net 1.5 GB · Apache-2.0
ChineseMarket
KnowLine Gen1
by PelaAI
Decisions trained mostly in Chinese.
Net 1.8 GB · Apache-2.0
GeneralMarket
Jeff
by jeff-legacy
Quick decisions in front of a big model; unsure queries go on to the big one.
Net 1.7 GB · Apache-2.0
GeneralMarket
Saina Helm 2
by Saina
Typed questions over any state: yes/no, single or multiple choice, ratings.
Net 1.7 GB · Apache-2.0
GeneralMarket
jev-decisions v1
by GhostScientist
Picks one of 2 to 24 labelled options from a structured state.
Net 1.5 GB · Apache-2.0
GeneralMarket
Bruv1
by bvolpato
Reads an evidence state and scores two to sixteen options; text only.
Net 1.5 GB · Apache-2.0
ScreensMarket
OneJev 1.0
by OmniJev
The first OneJev: screenshots, photos, video and text from real agent runs.
Net 2.2 GB · Apache-2.0
SafetyMarket
HaloGuard 0.8B
by astroware
Flags unsafe prompts and names the policy before they reach an LLM.
Net 2.2 GB · Apache-2.0
SafetyMarket
ImageShield
by prithivMLmods
Content-safety calls on images.
Net 1.7 GB · Apache-2.0
IT and securityMarket
Decision WAF 1.1
by 0xBrandon
A web firewall: reads an HTTP request and calls it safe or an attack.
Net 1.7 GB · Apache-2.0
FinanceMarket
Invoice triage
by distil labs
Sorts an accounts-payable inbox: invoice, receipt, reminder, vendor or spam.
Net 1.5 GB · Apache-2.0
ChineseMarket
Chinese Decision
by quincyqiang
Typed decisions in Chinese.
Net 1.5 GB · Apache-2.0
ArabicMarket
Ara-JEV
by AhmedReda25300
A compact Arabic decision model.
Net 1.7 GB · Apache-2.0
TurkishMarket
TR Telekom Router
by eqlx
Routes Turkish customer messages to one of six teams, with a confidence.
Net 1.8 GB · Apache-2.0
BaseBase
Qwen3.5 0.8B
by Qwen
The plain base every lite body grew from. A spoonful pulls a soup back toward it.
Net 1.8 GB · Apache-2.0
Good to know
- Same family only. Models built on different bases have different shapes, so their weights can’t be averaged. Each family here shares one Qwen3.5 base.
- The first model’s files are kept. Its config, tokenizer and answer layer come with the merge. Put d3-lite or d3-mini first to keep the z3 answer layer.
- Nothing is uploaded or scored. The merge runs locally and stays on your machine. A merge can beat every model in it or fall behind them; to measure yours, enter it in a pool. A mix of house models and listed public models earns a reduced rate there.
- Check every licence. Every model listed is Apache-2.0. Each model’s licence is copied into the folder, and
merge.jsonrecords the recipe.