Skip to content
Nuro AI Labs
Try HypersaveRead AVALON-2B
Research

Open foundation research on the path to general intelligence.

We publish weights, training code and synthetic-data pipelines wherever the research permits. Apache 2.0 is the default. We don't believe the systems on the path to general intelligence should be closed boxes held by three companies.

TeamsSelf-RAGMemory architecturesReasoningOpen Science
Featured paper
Released · April 2026Apache 2.0Sub-3B · Self-RAG

AVALON-2B

The first sub-3B language model that knows what it doesn't know.

1.88B parameters built on Qwen 3.5 2B. A five-token reflection vocabulary — [Retrieval] [No Retrieval] [Relevant] [Utility:5] — and a 22M-parameter MiniLM router at 90.5% accuracy. 82.5% Self-RAG token accuracy under LoRA, 40 tok/s on Apple M3, 12 tok/s on iPhone 15 Pro.

Akhil Ponnada · Naga Sri Arvapalli · Nuro AI Labs
Read the paper →Hugging Face ↗Ollama ↗
REFLECT architecture · live
router 90.5%
›What's in the news today about lithium prices?
router · MiniLM
22M params · 5ms latency
decides:
[Retrieval]
generation · Qwen 3.5 2B + LoRA
1.88B · 18 GDN + 6 softmax
emits:
[Utility:5]
More research

Two more lines on the road from personal to general intelligence.

Preprint imminent

PLMR

Pre-tokenizer Latent Memory Routing for byte-level language models.

A new design slot in byte-level language modeling. External persistent memory consulted before patch boundaries are placed, with retrieval similarity modulating the entropy threshold itself. Validated at toy scale with 7–10% relative loss reduction on evidence-bearing windows. Phase 1 BLT-1B continual-pretraining is underway.

Byte-levelMethodologyBLT extension
Akhil PonnadaRead more →
In active development

Hydra

A next-generation model architecture in active development.

We don't pre-announce. Hydra is the third research line on the road from personal intelligence to general intelligence. More on release.

ArchitectureIn flight
Nuro AI LabsRead more →
Publications

Everything the lab has put out.

Papers, preprints and technical notes. Open-source by default — Apache 2.0, weights, code and data on Hugging Face.
Apr 2026
Self-RAG
AVALON-2B — the first sub-3B Self-Reflective RAG language model
Akhil Ponnada · Naga Sri Arvapalli
Read →
2026 · preprint
Byte-level
PLMR — Pre-tokenizer Latent Memory Routing
Akhil Ponnada
Read →
Open by default

Apache 2.0 is the default for our research output. AVALON-2B ships with weights, GGUF quants and the synthetic-data recipe on Hugging Face and Ollama. PLMR and Hydra inherit the same posture. We don't believe the systems on the path to general intelligence should be held inside three companies.

Hiring

Come build the next paper with us. Research engineers, applied engineers, GTM.

We hire on a rolling basis. The bar is the work — show us a paper, a model, a system you shipped. The lab is small, the surface area is large, and every hire moves a research line.

Join the Research team →careers@nuroailabs.com