World’s only company with six Small Language Models focused on education at scale — from chat generation to vision AI. Each one distilled from large LLMs to be privacy-focused, classroom-safe and tailored for education.
Why Sumeru exists
Cloud LLMs were never designed for a world classroom. They charge per token, need WiFi, send student data to US servers, and don’t speak our languages well. We built six small models that solve all of that.
Schools worldwide love the idea of AI tutors but can't afford per-token cloud bills, don't have reliable internet, and won't risk uploading student data to foreign clouds. The biggest blocker to AI in education isn't capability — it's the runtime.
Instead of training another giant LLM, we distilled six narrow ones from the best frontier models. Each one keeps the accuracy of its larger teacher on its specific education job, but fits in 100 MB–2 GB.
Every Sumeru model is trained on curriculum-aligned data, audited by teachers, guardrailed for age-appropriateness, and tuned for 32+ world languages. Not generic with an edu skin — purpose-built.
From 100 MB to 2 B parameters. Four available today; two on the roadmap.
Audio-native intelligence: voice interaction, sound identification and real-time transcription — all on-device.
Register NowStep-by-step reasoning for plans, breakdowns and structured analysis — without sending the chain-of-thought to a cloud.
Register NowVision-based understanding — diagram parsing, OCR and handwritten work recognition. World's first vision AI for the classroom.
Register NowThe everyday workhorse. General language understanding across domains — fast, classroom-safe, near-zero hallucination.
Register NowDeep research and long-context analysis for knowledge-heavy workflows. The largest model in the suite, distilled for the edge.
On the roadmapThe lightest, fastest model. Tuned for short queries and embedded workflows — runs everywhere the larger models can't.
On the roadmapPicking your model
Six small models, six education-shaped jobs. Find your match below.
Voice apps, transcription, sound ID
Audio-native · real-time · on-device
Step-by-step reasoning, math, plans
Chain-of-thought stays on the device
Diagram, handwriting, document parsing
World's first vision AI for classrooms
General edu chat & content generation
The everyday workhorse
Deep research & long-context analysis
Heaviest model, distilled for the edge
Embedded short queries, kiosks, IoT
Runs where the larger models can't
Not sure which fits? Ask sumeru@schoolexl.com — we’ll recommend in a paragraph, not a sales pitch.
Trained against the best frontier models, then compressed for the edge — keeping the accuracy, losing the bulk.
Runs on-device or inside your network. Student data, classroom audio and exam answers never leave the building.
Strict guardrails on toxicity, hallucination and age-appropriateness — the only education AI trained and approved by teachers.
Six small models, each tuned for an education-shaped job: chat, audio, vision, reasoning, research. No generic compromise.
Under the hood
Stage 01
We start with the strongest open frontier models — Llama, Mistral, Phi. They're the teachers from which our students learn.
Stage 02
Each Sumeru model is distilled to a single education job through a teacher-student pipeline. Narrow accuracy beats general intelligence.
Stage 03
Trained on world curriculum-aligned data, then audited by working teachers across grade levels. Hallucination, toxicity and age-fit are graded — not assumed.
Stage 04
Compiled to run on Intel iGPU/NPU, NVIDIA GPUs, ARM CPUs and Snapdragon NPUs. Zero cloud calls at inference.
Runs everywhere your users are
Performance proof
On the narrow education jobs Sumeru is trained for, on-device accuracy is within 6% of GPT-4o — at 1/2000 the per-query cost and zero data exfiltration.
Generic SLMs are fine-tuned by the whole internet. Sumeru is fine-tuned by teachers — on the actual textbooks, syllabi and language varieties students learn from.
Numbers illustrative · full methodology in our technical paper, on request
Three ways to ship Sumeru
From a one-day evaluation to a custom fine-tune on your private data, there’s a path that fits your team and your timeline.
Hosted access to the four live models.
Best for: Researchers · prototypes
Request sandbox accessSumeru shipped on a workstation — managed for you.
Best for: Schools · branches · enterprises
Scope a deploymentA Sumeru variant trained on your private dataset.
Best for: Edtech · publishers · partners
Talk to research teamfrom sumeru import SumeruI
# Load on-device · no network call
model = SumeruI(device="cuda")
# Curriculum-aware tutoring in Hindi
answer = model.tutor(
"Explain photosynthesis to a 5th grader.",
language="hi",
grade=5,
)
print(answer)Illustrative. SDK and docs ship with sandbox access.
Not at the moment. The four live models are accessible via our hosted sandbox and through EdgeAI deployments. We're evaluating an open-weights release for the smaller models — get in touch if your use case depends on it.
Something not covered? Email sumeru@schoolexl.com — your reply lands in our research team’s inbox.
Tell us what you’re building and which Sumeru model fits. We’ll get you trial access for evaluation, or scope a deployment.