What is the sarvam-jev project?

An open, generation-free typed-decision engine on Sarvam's Indic models from SAGAR-TAMANG: choices are lettered slots, the probability distribution is read from those letters' logits, and no text is ever sampled.

How does sarvam-jev relate to Jev?

The README says the interface pattern is TypeSafe's Jev — one state, many typed questions, answers as probabilities — and it measures itself against openjev's published ladder rather than claiming wire compatibility with TypeSafe.

Why build this on a Sarvam model?

Tokenizer economics: the README says sarvam-1 encodes Devanagari at roughly 4 characters per token where Qwen-class tokenizers are 2–4x worse, which matters when the whole argument is paying for a state once and then asking many cheap questions.

What do the Phase 0 results show?

On 144 authored decisions with 3 options each, sarvam-1 2B base scored 0.516 mean family balanced accuracy against 0.333 chance; published openjev numbers for larger models include MiniCPM5-2B at 0.686 and Qwen3.5-4B at 0.813.

Does better prompting close the accuracy gap?

The README says no: 0 shots scored 0.490, 1 shot 0.470, 3 shots 0.516, and the chat template 0.404. With 99.8% of probability mass on answer slots, the format is learned almost perfectly while decision semantics are what a 2B base model lacks.

What hardware does it need?

Phase 0 ran on a single RTX 3060 with sarvam-1 in BF16. The project description says the engine runs client-side in the browser; the digest shows no separate inference hardware requirements and no install steps.

What is planned next, and is it licensed?

Phases 1–3: an Indic decision fixture with a sarvam-1 versus Qwen head-to-head, packed branch attention over one shared KV, then a LoRA calibration head reporting ECE. GitHub metadata lists no license, and the repo was created and last pushed on 2026-09-18.

Jev directory

CommunityJev Models & Reimplementations

Sarvam Jev

A generation-free typed-decision engine on Sarvam's Indic models: options become lettered slots, the answer is read from those letters' logits, and one state prefill is amortised across many cheap questions. Its README says the interface pattern is TypeSafe's Jev.

See highlights Open github.com

Stars: 60
Language: Python
Reading time: 3 min
Published by: Community
Author: SAGAR-TAMANG
Last verified: Oct 2026

sagar-tamang/sarvam-jev

User repository on GitHub

View on GitHub

Generation-free typed decisions on Indic LLMs. An open Jev-style inference engine on sarvam-1: constrained logit readout instead of autoregressive JSON. Runs client-side in the browser.

Forks: 10
Last pushed: Updated Sep 2026

Repo stats from the GitHub API, cached Oct 2026.Homepage

Tagscommunity github python indic inference

Works forJev for Classification Jev for Scoring

Used inJev for Software & DevTools

Languages: Python
Platforms: Browser
Capabilities: Open sourceRuns locally

Highlights

README: options become lettered slots and the distribution is taken over the token IDs of those letters, so nothing is sampled and nothing can be malformed.
The interface pattern is TypeSafe's Jev — one state, many typed questions — and one expensive state prefill is amortised across every question asked about it.
Phase 0 ran 144 authored decisions with 3 options each on one RTX 3060: sarvam-1 2B base scored 0.516 mean family balanced accuracy versus a 0.333 chance line.
Shared-state path on the Hindi preset: five criteria over one ticket took 1489 ms fresh versus 280 ms shared — 5.31x — with prompt tokens down from 2521 to 737.
99.8% of probability mass landed on answer-letter slots, but the README says decision semantics are missing, so the Phase 3 calibration head is required.

Watch out

README reviewed 2026-10-17 and truncated mid-sentence in the shared-state caveat, so install commands, the rest of the Phase 0 evidence, and a license (the metadata lists none) are not in the digest. Accuracy and timing numbers come from one RTX 3060 running sarvam-1 in BF16 and are the project's own, measured against published openjev figures rather than a shared evaluation run.

Questions about this entry

Short answers drawn from the linked project or page.

From the Jev guides

Original write-ups on what this entry is for.

Beyond text and boxes: multimodal decisions you can run yourself

Oct 16, 2026

Jev-Omni's merged classifier scores text, image, audio, and video options; VideoJev and Laya Vision show what open multimodal decision models accept, what they return, and which numbers are the authors' own.

Jev Models & Reimplementations#multimodal#open-models#vision

Read article

Jev 1.14: Images, Box, and the Danger Zone

Oct 15, 2026

Jev Day shipped a routine model bump and a deliberate widening of the API: images on both models, a box question type on 1.14 only, and 262k context, all behind danger-zone headers.

Jev Models & Reimplementations#release#vision#danger-zone

Read article

The Open Decision-Model Roundup

Oct 14, 2026

Training recipes, looped architectures, and language ports: what the open side of the typed-decision class shipped this month, and how to pick by runtime and calibration instead of the leaderboard.

Jev Models & Reimplementations#open-models#training#local

Read article

From the Jev AI model community

Posts from builders shipping with Jev right now.

Follow @typesafeai

Sarvam Jev

Highlights

Questions about this entry

What is the sarvam-jev project?

How does sarvam-jev relate to Jev?

Why build this on a Sarvam model?

What do the Phase 0 results show?

Does better prompting close the accuracy gap?

What hardware does it need?

What is planned next, and is it licensed?

More Jev resources like this

From the Jev guides

From the Jev AI model community