2024
Incorporated
Nuro AI Labs Limited registered in London (Companies House #16079959). The thesis: personal intelligence as the bottom-up route to general intelligence.
NewIntroducing AVALON-2B — the first sub-3B Self-RAG language model.Introducing AVALON-2B→
Nuro AI Labs · 2026London · Companies House #16079959
Building personal and general intelligence.
An AI research lab. Personal intelligence is the through-line — but we work across language models, architectures, reasoning, agents and the systems on top.
We don't pre-announce. We ship.
Now live · AVALON-2B · Hypersave · Khyaa
Scroll ↓Supported by
Featured · Research
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. A 22M-parameter MiniLM router at 90.5% accuracy. 82.5% Self-RAG token accuracy under LoRA fine-tune. Apache 2.0. Live on Hugging Face and Ollama.
[Retrieval][NoRetrieval][Utility][Relevant][Continue]
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]Latest from the lab
In production, in flight, and in development.
How we work · three concentric layers
Research → infrastructure → applied.
Same shape as Anthropic, OpenAI and DeepMind. Open-weights at the research layer, commercial at the product layer. The applied work isn't a distraction from the research — it's the test of it.
01Open by default
Research
Apache 2.0 papers and weights. AVALON-2B is live; PLMR preprint imminent; Hydra in active development.
AVALON-2BPLMRHydra
02Commercial
Infrastructure
The cognitive memory layer for AI agents. SOC 2 Type II. Sub-200ms p95. Beats every adjacent system on LoCoMo.
Hypersave
03Commercial
Applied
Vertical and consumer products built on the same stack. Compliance-aware, citation-backed, on-device first.
KhyaaNuro StudioNuro ChatNuro One
Thesis · 01
The path to general intelligence runs through personal intelligence.
We're a research lab — first and foremost. That means we work on the open questions across language models, architectures, reasoning, agents and the systems on top of them, and we publish what we learn.
Personal intelligence is the through-line that ties our work together — persistent memory, self-reflection, per-agent cognition. It's the part of intelligence the current frontier mostly skips, and we think it's a credible bottom-up route to the general intelligence everyone is racing toward top-down.
But personal intelligence isn't the only thing we work on. Some of our research is on language-model reflection (AVALON-2B). Some is on byte-level architectures and routing (PLMR). Some is on systems we haven't named yet (Hydra). Some ships as open weights; some ships as commercial infrastructure (Hypersave) or applied products (Khyaa, the Nuro stack). — one lab, many bets.
agent.pypython
from hypersave import Memory
mem = Memory(agent_id="researcher")
# every turn, the agent both remembers and reflects
mem.write("Sarah's youngest is heading to Stanford in August.",
sector="episodic")
ctx = mem.recall("What's top of mind for Sarah right now?")
# → cited, ranked, decay-weighted answer — not chunksHypersave SDK · TypeScript and Python · npm i @hypersave/sdk
Flagship infrastructure
Memory that thinks.
Hypersave is the cognitive memory layer for AI agents. Brain-inspired architecture: five cognitive sectors, each with its own Ebbinghaus decay curve. Knowledge graph + vector + keyword + RRF hybrid retrieval. Ten-stage query pipeline. Answer synthesis — not chunks.
LoCoMo
86%
p95 latency
<200ms
Token savings
92%
Fact accuracy
94%
Cognitive sectors · Ebbinghaus decay
S(t) = S₀ · e-λt
0d700d
LoCoMo · long-context memory benchmark
86%
Hypersave v1 · audited March 2026
Reported in good faith from internal benchmarks. Reproducible script will ship with the v1.1 release notes.
Sample brief · cited
14:45 → 15:00 meetingSarah Chen joins at 15:00. Last note flagged inflation; today's CPI print came in 30 bps below consensus[1]. Recent touchpoints: Roth conversion (Oct 12), Stripe RSU vesting (Oct 4), 529 contribution limits (Sep 21)[2]. Open commitments: Q3 statement due Friday; estate-attorney intro by Nov 1[3].
[1] CRM note · 2026-03-12
[2] Meeting transcripts · Sep–Oct
[3] Voice capture · 2026-04-22
Applied · Khyaa
Never walk into a client meeting cold.
Personal intelligence for US financial advisors. Pre-meeting briefs, on-device voice capture, natural-language Ask across your entire book — every answer cites its source.
- 200-word brief 15 minutes before each external meeting
- 1-button voice recorder — recap leaves your phone as text, not audio
- Ask the book in English; every claim has a source
- SEC marketing rule, Reg BI, Rule 204-2 retention aware
Products · the personal-intelligence stack
Built on the same research stack.
Principles · how we operate
Seven things we hold ourselves to.
Stated up front so the next time you read a Nuro release, you can hold us to them.
and so we don't lose the plot01Open research, commercial products.
Open research, commercial products.
Our research output is open by default — weights, training code, and the synthetic-data pipelines that produced them. AVALON-2B is the proof. Our products are commercial: Hypersave, Khyaa and the Nuro stack are how we fund the lab. The split is deliberate, and it's how Anthropic, OpenAI, Mistral and DeepMind all actually operate.
02Personal first.
Personal first.
Every advance we ship has to make at least one specific mind better — one user, one agent, one team — before it makes the world better. We do not chase abstract benchmarks at the cost of the unit we serve.
03Compounding, not theatrical.
Compounding, not theatrical.
We ship in public, incrementally. We do not pre-announce. We do not promise products we have not built. The published research is the marketing.
04Accountable.
Accountable.
Every claim is auditable. Every benchmark we publish has reproducible code. Every Khyaa answer cites its source. Every Hypersave query carries a confidence score. We treat citation as a first-class product feature.
05Useful before grand.
Useful before grand.
A frontier lab that doesn't ship what it builds is a research department. We ship Hypersave, Khyaa and Nuro Studio alongside AVALON and PLMR. The applied work is not a distraction from the research — it is the test of it.
06For every mind.
For every mind.
Personal intelligence is not luxury infrastructure. The cognitive memory layer should run on a phone, in a regulated bank, in an SME, in a classroom — not just in a hyperscaler. Geography of compute should not gate access to intelligence.
07Patient about the destination.
Patient about the destination.
General intelligence is the goal. We don't think it's three months away, and we don't think the path runs through a single 10-trillion-parameter pretraining run. We think it runs through compounding — through architecture, through memory, through reflection. We are willing to take a decade.
The labs that ship across the whole stack — open research, infrastructure, applied products — define the next decade of AI. The ones that pick one and stop don't.
By the numbers · live
3
Research lines
AVALON · PLMR · Hydra
7
Shipped products
infra + applied + open-weights
86%
LoCoMo (Hypersave)
head-to-head SOTA
1.88B
AVALON-2B params
Apache 2.0 · on-device
Our journey · a short version
From an idea about memory to a shipping research lab.
Four chapters, in order. The longer one is below.
2025
Hypersave research
Brain-inspired memory architecture — five cognitive sectors, Ebbinghaus decay, hybrid retrieval, ten-stage query pipeline, answer synthesis.
How Hypersave works →Q1 2026
Hypersave v1.0 GA
SOC 2 Type II. TypeScript + Python SDKs. 86% LoCoMo — beats every adjacent agent-memory system head-to-head.
Apr 2026
AVALON-2B released
The first sub-3B Self-RAG language model. Apache 2.0. Live on Hugging Face and Ollama.
Read the paper →Trajectory · in public
Frontier-grade work, incrementally.
We don't pre-announce. Every milestone is a thing that shipped.
- Nov 2024
Nuro AI Labs incorporated
Companies House #16079959. London-registered private limited. - Q1 2025
Hypersave research kicks off
Brain-inspired cognitive memory architecture for AI agents — five sectors, Ebbinghaus decay curves, hybrid retrieval. - Q3 2025
Khyaa first design partners
Personal intelligence for US financial advisors. Free beta with named RIA design partners. - Q1 2026
Hypersave v1.0 GA
SOC 2 Type II. TypeScript + Python SDKs. 86% LoCoMo — beats every adjacent agent-memory system head-to-head. - Apr 2026
AVALON-2B released
First sub-3B Self-RAG language model. Apache 2.0. Live on Hugging Face and Ollama. - Next
PLMR preprint · AVALON-3 · Hydra
Continued open research. Continued commercial product. Same compounding.
Team · the lab
A small lab. High signal.
Founded November 2024. We hire research engineers, applied engineers and GTM operators on rolling basis.
AP
01 / 03Akhil Ponnada
Founder & CEO
Author on AVALON-2B and PLMR. Sets research direction and overall company strategy. MSc International Business Management, Heriot-Watt University; BBA, Amity University.
NA
02 / 03Naga Sri Arvapalli
CTO
Co-author on AVALON-2B; led the AVALON training pipeline. Owns the technical architecture across research, Hypersave infrastructure and the applied product stack.
NY
03 / 03Naveen Yelloji
Executive Director
Senior CXO-level operator with three decades of leadership across AI, media & entertainment, telecom, infrastructure and technology. IIM Ahmedabad alumnus; University of Hull.
Hiring · 01
Research
Pretraining, fine-tuning, evals. AVALON, PLMR, Hydra. Open by default.
Open roles →Hiring · 02
Infrastructure
Hypersave platform engineering. Distributed memory, latency at p95, SOC 2.
Open roles →Hiring · 03
Applied
Khyaa, Nuro Studio, Nuro Chat, Nuro One. Vertical and consumer surfaces.
Open roles →Hiring · 04
Operations
GTM, partnerships, finance. Help us turn frontier research into a durable lab.
Open roles →Where we live
London. Quietly, deliberately.
Registered office at 128 City Road. Companies House #16079959. We're hiring across research, infrastructure and applied — remote-friendly, London-anchored.
Registered office
London
128 City Road, EC1V 2NX · United Kingdom
Founded
Nov 2024
Nuro AI Labs Limited · #16079959
Posture
Open + Commercial
Research is open by default. Products fund the lab.
Press
press@
press@nuroailabs.com
Press inquiries
press@nuroailabs.comOther inquiries
hello@nuroailabs.comPress kit
Boilerplate · bios · logos →What does "personal intelligence" actually mean?
Persistent memory + self-reflection + per-agent cognition. The set of capabilities a current frontier LLM does not have because it is restarted from scratch every conversation. Hypersave is the cognitive memory layer that adds those capabilities to any agent. AVALON-2B is the first runtime that knows when it doesn't know.
How is this different from OpenAI, Anthropic or DeepMind?
They are scaling top-down — bigger models, more compute, hoping general intelligence emerges from sheer parameter count. We are building bottom-up — the memory and reflection systems that compound personal intelligence into general intelligence. Both paths matter. Ours is under-served.
Are your products open source?
Our research is open. AVALON-2B is Apache 2.0; weights, GGUF quants and the paper are public. PLMR and Hydra inherit the posture. Our products are commercial — Hypersave, Khyaa and the Nuro stack are how we fund the research. The split is deliberate.
Where are you based?
London. Companies House #16079959. Registered office at 128 City Road, EC1V 2NX.
Who is the team?
Akhil Ponnada (Founder & CEO). Naga Sri Arvapalli (CTO, AVALON-2B co-author). Naveen Yelloji (Executive Director). We hire research engineers, applied engineers and GTM operators on rolling basis.
Are you raising?
We are talking to a small set of seed investors who share the personal-intelligence thesis. Reach press@nuroailabs.com.
Newsletter · slow channel
One email per release. That's it.
Papers, weights, and product updates as they ship. No marketing, no tracking, unsubscribe anytime.
Newsletter
Slow updates from the lab.
One email per release. No marketing, no tracking, unsubscribe anytime.
Personal intelligence. General intelligence.
Build with us. Read the research. Try the products. Or just say hello.
nuroailabs.com·@nurolabs·huggingface.co/nuroai