Project-Labs-x
ProjectLabs-x

Project-Labs-x / Frontier Language Model

HEKA

A frontier language model built for reasoning, code, and long-context tasks. Designed to be honest about what it doesn't know.

Join the waitlistRead the research
SCROLL
128K
Token context
40+
Languages
88.4
MMLU score
2026
Release target

Capabilities

Built for real tasks.

⟨/⟩

Code generation

Writes, debugs, and explains code across 40+ languages. Understands context across large codebases.

Long-context reasoning

128K context window. Reads, summarises, and cross-references entire documents and conversations.

Structured output

Reliable JSON, markdown, and typed schema output without wrapper prompts or post-processing.

Instruction following

Follows complex, multi-step instructions with high fidelity. Clarifies ambiguity rather than guessing.

Tool use

Function calling with parallel tool execution. Designed to operate reliably in agentic pipelines.

Δ

Multilingual

Strong performance across major European and Asian languages. Not just English-first fine-tuned.

Benchmarks

Scores, not claims.

MMLU88.4
HellaSwag94.1
ARC-Challenge91.7
WinoGrande87.3

Methodology note

All benchmarks are run zero-shot unless stated. We do not tune on benchmark test sets. Numbers are preliminary — Heka is still in training. Final scores will be published alongside model weights at launch.

Status

Training run active — ETA Q1 2026