Tara Canvas
@taraspace/canvas-core
Portable document contracts, geometry, ontology operations, React canvas primitives, and standalone renderers.
Open source software
High-performance inference and spatial systems.
We build and open source foundational inference engines, spatial canvas runtimes, and developer frameworks for serious intelligence and data workflows.
5 live projects·3 developer tools·Rust & TypeScript
Available now · Open source
Taraference (tarafer)
Hyper-optimized single-user LLM & embedding inference engine built in Rust — custom-engineered for maximum single-stream performance and low latency.
LanguageRust (100%) + Custom CUDA Kernels
Target HardwareAccelerated compute systems
SpecializationSingle-stream zero-overhead inference
Repositoryagkomyint/taraference
Rust & Custom CUDA Kernels
Bypasses generic Python runtime layers with custom hand-tuned CUDA kernels and strict zero-allocation Rust memory pipelines.
Single-Stream Optimization
Engineered specifically for dedicated single-user workstations and latency-critical prompt ingestion without queue contention.
Model Family Tuning
Hardware-level KV-cache layouts, customized quantization kernels, and optimized tensor ops tailored for specific model architectures.
LLM & Dense Embeddings
Unified high-speed inference for both generative language models and dense spatial/semantic embedding indexing.
taraference-quickstart.shbash
# 1. Clone the repository
$ git clone https://github.com/agkomyint/taraference.git
$ cd taraference
# 2. Build with CUDA acceleration
$ cargo build --release --features cuda
# 3. Launch single-stream inference engine
$ ./target/release/tarafer --model qwen-2.5-7b --device cuda:0
Available now · MIT licensed
TaraBG
A clean-room Rust implementation of BGZF, designed for interoperable genomic compression, indexing, integrity checks, and random-access workflows.
41/41
tests passing
1.40x
faster in the audited VCF run
9,747
identical queried records
HTSlib validated, identified, indexed, and queried TaraBG output from a real 99.2 MB 1000 Genomes VCF slice. Performance varies by data, level, threads, and hardware.
Open Weights & Datasets · Hugging Face
Tara Model Series
Custom-trained language model series optimized for regional intelligence, spatial ontology reasoning, multi-step evidence synthesis, and structured query parsing. Available openly on Hugging Face.
Current ReleaseTara 1.4.1 (Text Generation)
Next GenerationTara 1.5 (In Active Training)
Lineage1.1 → 1.2-quest → 1.3 → 1.4-base → 1.4.1
Publisherhuggingface.co/aungkomyint
v1.4.1 · AvailableHF Repo
Tara 1.4.1
Fine-tuned instruction model specialized in structured entity extraction, regional question answering, and synthesis over complex investigation documents.
text-generationtara1.4-base
v1.5 · In DevelopmentNext Release
Tara 1.5
Next-generation architecture featuring expanded context reasoning, direct spatial canvas tool calling, and high-precision knowledge graph alignment.
extended contextgraph reasoning
Open Training Datasets
Includes risk-question-taxonomy-30k, a 30,000-entry dataset for fine-tuning question decomposition and risk evaluation.
Explore 30k Taxonomy DatasetNative Engine Acceleration
The Tara model series is natively supported by the Taraference engine with zero-overhead single-stream inference and custom CUDA execution kernels.
tarafer --model tara1.4.1quickstart_tara.pypython
# Load Tara 1.4.1 directly from Hugging Face
from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = "aungkomyint/tara1.4.1"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto")
Released September 2026
The Tara developer stack is open source.
Build a canvas locally, compile it from files, or control a hosted workspace from the terminal. Each layer is MIT licensed and independently installable.
Bare-metal performance
Core compute engines written in Rust and CUDA with direct memory control and custom kernels, eliminating multi-layered runtime overhead.
Inspectable by default
Open source foundations ensure your AI pipelines, spatial documents, and inference workflows remain fully verifiable and free from vendor lock-in.
Modular architecture
Each project solves one focused layer—inference, spatial canvas runtime, or application framework—with strict contracts and clear separation.
Build, inspect, and contribute.
Explore the Taraference repository or follow our roadmap as more modules become open source.