APN.

Research Lead · Tether / QVAC

I work on efficient language models for on-device use.

My work covers pretraining, post-training, quantization, runtime integration, and evaluation across mobile, desktop, and server hardware.

Research Public releases

Selected work

148B
synthetic educational tokens released in Genesis I & II
2T
tokens used in distributed pretraining
5
public QVAC releases as a contributor
13B
largest BitNet model fine-tuned on an iPhone 16

This work includes multi-node H100 training and evaluation across mobile, desktop, and edge GPU hardware.

Research focus

Efficient models, from training to deployment.

I work across data, training, quantization, runtime integration, and evaluation. The objective is to build models that remain useful while running privately on devices with limited compute and memory.

01

Efficient model research

Low-bit and ternary language models for local inference under memory and compute constraints.

02

Edge fine-tuning

LoRA fine-tuning across Vulkan, Metal, consumer GPUs, and mobile devices.

03

Training and evaluation

Data generation, pretraining, SFT, distillation, reinforcement learning, and model evaluation.

04

Research leadership

Experiment design, technical review, hardware evaluation, publication, and coordination with engineering teams.

Research and releases

Selected public work

These are team projects. Each entry summarises the work and states my contribution.

Data · October 2025

Read release

QVAC Genesis I

A synthetic educational dataset generated from model failures across nine domains.

My contribution: I built and operated data-generation and training infrastructure, ran pretraining ablations, and contributed to the release analysis.

  • 41B public synthetic tokens
  • 9 educational domains
  • Downstream and model-based evaluation

Data · December 2025

Read release

QVAC Genesis II

An expansion of Genesis with option-level reasoning, ten additional domains, and 107B new tokens.

My contribution: I worked on the data and evaluation pipelines, distributed training, ablations, and release analysis.

  • 148B tokens combined with Genesis I
  • 19 educational domains
  • 107B new tokens across 86M samples

Edge training · December 2025

Read release

QVAC Fabric LoRA Fine-Tuning

A LoRA fine-tuning workflow for desktop and mobile GPUs using Vulkan and Metal.

My contribution: I built the original proof of concept, curated the dataset, designed experiments, and validated Vulkan results against PyTorch.

  • Hardware-agnostic workflow
  • Mobile-GPU fine-tuning
  • Qwen3 and Gemma 3 support

Low-bit systems · March 2026

Read release

BitNet b1.58 Edge Fine-Tuning

On-device fine-tuning and inference for ternary language models on mobile and consumer GPUs.

My contribution: I benchmarked the low-bit path, reviewed implementations, evaluated quality, tested mobile and desktop hardware, and contributed to the paper.

  • Up to 13B parameters fine-tuned on an iPhone 16
  • 1B model fine-tuned on a Samsung S25
  • 2.1–11.3× faster edge-GPU inference than CPU

Health AI · May 2026

Read release

QVAC MedPsy

Medical language models at 1.7B and 4B parameters, with quantized variants for local deployment.

My contribution: I contributed to the research and evaluation, helped prepare the public release, and presented the work in Shanghai.

  • 1.7B and 4B model releases
  • Quantization studied for mobile deployment
  • Seven closed-ended medical benchmarks

Current research

In progress

Dense-to-Ternary Language Models

Current research on converting dense language models to ternary weights.

Current scope: Conversion methods, training stability, distillation, runtime integration, and on-device evaluation.

Benchmarks, training recipes, and artifacts will be added after the work is complete.

Models & datasets

Public artifacts

Public checkpoints, datasets, and adapters associated with the releases above. Each entry includes repository metadata and my contribution.

Areas of contribution

  1. Data creation
  2. Pretraining
  3. Post-training
  4. Quantization
  5. Runtime integration
  6. Mobile evaluation
  7. Publication

Loading the latest public snapshot…

Systems tested

  • H100 clusters
  • AMD GPUs
  • Apple Silicon
  • Snapdragon & Mali
  • iPhone-class devices
  • Vulkan & Metal
  • CPU & llama.cpp

Evaluation

Evaluation across models, runtimes, and hardware

I compare real and simulated quantization, inspect loss curves, check numerical agreement across runtimes, measure perplexity and benchmark retention, and test throughput and memory use on devices.

Attribution

Clear contribution statements

These releases were produced by research and engineering teams. Each card states the areas where I contributed.

Public work

Publications and profiles

01
Hugging Face author profileTechnical articles and public releases
 02
QVAC on Hugging FaceModels, datasets, and technical reports
03
Current researchManuscripts will be added when they are public or formally accepted.

Talks

21 May 2026 · Shanghai

QVAC MedPsy: Medical and Healthcare Language Models for Edge Devices

I presented MedPsy at muShanghai, covering model development, evaluation, quantization, and deployment on edge devices.

Event details

Experience

Research and engineering experience

My experience spans edge speech systems, production LLM applications, distributed model training, synthetic data, and low-bit model research.

Feb 2026 — present

Research Lead · Tether / QVAC

I lead research on efficient models and edge fine-tuning, including experiment design, evaluation, hardware testing, and publication.

May 2024 — Feb 2026

AI Engineer, LLM · Tether

I worked on large-scale pretraining, synthetic data, post-training, quantization, and distributed training infrastructure.

Oct 2023 — Apr 2024

AI Implementation Engineer · Peartree

I built edge speech and interview-assessment systems, including Whisper experiments on mobile GPUs.

Oct 2022 — Oct 2023

Software Engineer, AI/ML · Gigaforce

I developed LLM applications for fine-tuning, retrieval, and document processing.

Akshay P Nambiar

Background

Engineering background

I completed a B.Tech in Civil Engineering at IIT Gandhinagar and later moved into robotics, edge speech, and language models. My work now combines model research with systems engineering.

I have also represented IIT Gandhinagar in badminton, received the Explorer’s Fellowship, and worked on robotics projects.