Open weights

Vector 1 — open-weight models, built to assist teachers.

Vector is MikaLabs' family of open-weight models built to assist teachers — not students. Post-trained to produce the materials educators make over and over: differentiated worksheets, mark schemes, misconception guides, lesson plans and pitched explanations across maths and the sciences. Apache 2.0, and small enough to run locally and offline.

Developer
MikaLabs
License
Apache 2.0
Language
English
Subjects
Maths · Biology · Chemistry · Physics
Runs
Locally & offline, on modest hardware
The family

Four models, one pedagogy.

A family released under a permissive license — weights, tokenizer, and model card in the open. Rolling out one model at a time.

Vector-L1-4B
4B params · language · Apache 2.0
Open weights
“Light, version 1” — the first and smallest Vector model. Compact enough to run locally and offline on modest consumer hardware, with quantized GGUF builds included.
Vector-1-14B
14B params · language · Apache 2.0
In training
The flagship, scaling the teaching-assistant recipe to 14B for deeper subject reasoning. Currently in training.
+In training — weights on the way
Vector-R1-14B
14B params · reasoning · Apache 2.0
Coming soon
A reasoning-tuned variant — extended chain-of-thought for multi-step problem solving and worked solutions.
+Coming soon
Vector-V1-8B
8B params · vision · Apache 2.0
Coming soon
A multimodal variant that reads diagrams, handwriting and worked solutions alongside text.
+Coming soon
What it does

Built to make classroom materials.

Vector is specialised for the resources teachers produce again and again — the work, not the chat.

ADifferentiated worksheets
Multi-tier support / core / extension sets with genuine difficulty progression — not the same questions relabelled.
BExam-style mark schemes
Method marks (M) and answer marks (A) shown separately, the way real exam marking actually works.
CMisconception guides
The specific, subject-level errors students make on a topic — each listed alongside how to address it.
DStructured lesson plans
Objectives, starters, main activities and plenaries, laid out ready to take straight into a classroom.
EMixed-format questions
Short answer, true/false with justification, fill-in-the-blank, calculation and explain-your-reasoning — never defaulting to multiple choice.
FHolds the brief
Follows complex, multi-part instructions, pitches difficulty to the age or ability you name, and returns clean, ready-to-use output.

Vector is a specialist, not a general chatbot — it isn't built for coding, open-ended chat, or subjects beyond maths and the sciences, and high-stakes assessment should always have a teacher reviewing the output. Full weights and model card live on Hugging Face.

MikaLabs

An advanced AI research lab building the learning systems universities trust to teach.

Frontiers
Research
© 2026 MikaLabs[email protected]