Vegaml

Vegaml is a machine learning library providing typed decisions from a frozen language model and a physics engine.

Vegaml combines a frozen multimodal language model with a small physics engine to generate direct decision outputs like choices, scores, and probabilities. It supports states up to 73,728 tokens with read-once prefix caching and a memory reader to process long documents without silent truncation. The system offers multiple answer types including choice, score, and boolean, each carrying calibrated probabilities and explicit abstain flags. It runs in two sizes, 0.8B and 4B parameters, using Qwen3.5 backbones. Users can fine-tune rank-32 adapters on custom data while keeping the base language model frozen.

Vegaml at a glance

Best for: Software engineer, Researcher
Runs on: Python, self-hosted
Pricing: Free
Released: October 2026
Category: AI research tool
Website: github.com/nandhakishorm/vegaml

Key features

Processes states up to 73,728 tokens
Supports image and text inputs
Provides calibrated probabilities and conformal answer sets
Includes explicit abstain flags
Offers choice, score, and boolean answer types
Features read-once prefix caching
Runs locally on your own hardware

Use it for

Intent classification
Document analysis over long texts
Automated decision making with calibrated confidence
Fine-tuning custom adapters on private datasets

Alternatives to Vegaml

NanoJev — It runs parallel decisions and outputs probability distributions without output-token decoding.
VLMEvalKit — Evaluate large vision-language models across numerous benchmarks with a single command.
Decido — Runs locally on CPU while beating published benchmark scores.
ZML — It enables Python-free inference while abstracting hardware complexities.
vLLM — Maximizes throughput with PagedAttention and continuous batching.
OpenLLM — It simplifies infrastructure management and optimizes inference speed and cost.
SemIf OpenJev — Extracts direct token probabilities without generating text or running JSON repair loops.
LMMs-Eval — Provides open research and tooling for multimodal model evaluation and benchmarks.
Paper2Agent — It enables direct interaction with research papers and automatic result reproduction.
Homi — It answers complex audio questions natively with high accuracy.
Vev — Runs locally on NVIDIA hardware with a fast, Jev-compatible API.
Lighteval — Evaluates LLMs across multiple backends with thousands of built-in tasks.

All Vegaml alternatives