Zapdev-labs

Zapdev-labs

Building faster LLMs on real hardware

Zapdev-labs builds quantization, inference runtimes, and benchmarks — C and Python tooling that cuts memory, raises tokens-per-second, and ships local /v1 chat HTTP APIs you can run on your hardware.

Contact meView projects

Current Focus

LLM Compression

Weight and KV-cache quantization — PolarQuant, FastVQ, and oxidize-quantize for edge deployment.

GitHubHugging Facezapdev.link

Local Inference

C and Python runtimes with local /v1 chat endpoints — oxidize, turboforge, and miniforge.

Open Source

Benchmarks, agent tooling, and inference stacks built in public for developers worldwide.

12+
Open-source projects
3-bit
KV cache compression
6×
KV memory reduction

Open Source Work

Tools and libraries built in public.

View all on GitHub
C

oxidize

A dependency-light C11 runtime for local LLM inference — native CLI, /v1 chat HTTP server, and CPU/GPU backends.

oxidize— stars
Python

turboquant

Near-optimal KV cache quantization with FastVQ — ~3-bit keys, ~6× smaller caches, GPU kernels.

turboquant— stars
Python

turboforge

Hardware-aware enterprise runtime with a llama.cpp backend plus benchmark telemetry.

turboforge— stars
Python

miniforge

High-performance MiniMax inference with GGUF quantization and TurboQuant KV modes.

miniforge— stars
Python

ollama-performance-benchmark

Back-to-back TPS benchmark on one GGUF — oxidize and miniforge vs stock llama.cpp.

ollama-performance-benchmark— stars
TypeScript

openagentmail

Self-hosted AgentMail-compatible email for AI agents — inboxes, threads, and webhooks.

openagentmail— stars

Selected Capabilities

We take on complex inference problems at the edge of what's possible and make them boringly reliable.

AI / INFERENCE

LLM infrastructure, quantization pipelines, and local serving stacks.

DEV TOOLS

Open-source libraries, hardware probes, and cross-backend benchmark harnesses.

PRODUCT / SYSTEMS

Interfaces and APIs that match the depth of the underlying inference system.

Recent Work

What we've been building lately

All projects
Open Source

oxidize

C11 local LLM inference runtime

Open Source

turboquant

KV cache quantization for LLM inference

Open Source

turboforge

Hardware-aware inference runtime

Open Source

miniforge

MiniMax serving with TurboQuant KV

Built With

llama.cppGGUFC11TritonConvexE2BHugging FacePyTorchMiniMaxOllamaoxidizellama.cppGGUFC11TritonConvexE2BHugging FacePyTorchMiniMaxOllamaoxidize

C rewrite

Oxidize is moving closer to the metal

The next Oxidize runtime is a C11 application: a dependency-light core, a native CLI, and explicit CPU and GPU backends. The rewrite is active in the public repository while releases and installers catch up.

Build the C runtime

$ git clone https://github.com/Zapdev-labs/oxidize.git && make -C oxidize/oxidize-c

Verify

Run the C test suite

make -C oxidize/oxidize-c test

CUDA

Build the GPU backend

make -C oxidize/oxidize-c cuda

The C rewrite is under active development. Inspect the source and current build targets: oxidize-c

How we work

Issue in, fix out

Most work starts on oxidize — open a GitHub issue, research the root cause, then use AI tooling to land a tested patch.

  1. 01

    Open an issue

    File a repro on oxidize — crash, wrong output, or slow inference

  2. 02

    Research

    Read the stack trace · profile on real hardware · trace root cause

  3. 03

    AI-assisted fix

    Cursor on the oxidize codebase · patch, test, open PR

Proof, not promises

Benchmarks you can rerun

ollama-performance-benchmark runs each backend sequentially on the same GGUF and prompt. On our hardware, oxidize and miniforge beat stock llama.cpp tokens-per-second.

View benchmark repo
ollama-performance-benchmark · results.csvtok/s
llama.cppsame GGUF
oxidizefastest
miniforgesame GGUF
Ollamasame GGUF

Back-to-back run on one prompt · Ryzen 7 PRO · Qwen3.5-4B Q4

Reference

Common questions

We are a machine learning lab focused on LLM performance: quantization (turboquant), local inference stacks (oxidize, miniforge, turboforge), and honest cross-backend benchmarks. We also ship agent infrastructure like openagentmail and product tooling like zapdev.

Stay Updated

Reach out for collaboration on quantization, local inference, or benchmark work.

jackson@email.zapdev.linkContact