Zapdev-labs

Faster language models on the hardware you own

Quantization, C11 inference runtimes, and benchmarks you can rerun. Built in public.

Three problems we keep pushing on

Compress the model, run it locally, then measure the result against the stock stack.

LLM compression

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

View repository

Local inference

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

View repository

Benchmarks

Back-to-back tokens-per-second runs on one GGUF, published so anyone can rerun them.

View repository

Every change is measured on real machines first

Open-source projects
12+Open-source projects
KV cache keys
3-bitKV cache keys
Smaller KV caches
6×Smaller KV caches

Open source, built in public

C

oxidize

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

View repository
TypeScript

openagentmail

Self-hosted, AgentMail-compatible email for AI agents. Inboxes, threads, and webhooks you run yourself.

C runtime

oxidize, rewritten closer to the metal

A dependency-light C11 core, a native CLI, and explicit CPU and GPU backends. The rewrite is active in the public repository while releases catch up.

How a fix lands

  1. Open an issue

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

  2. Research

    Read the stack trace, profile on real hardware, trace the root cause.

  3. AI-assisted fix

    Patch the oxidize codebase with Cursor, run the C tests, open a PR.

Build from source

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

Run the C test suite

$ make -C oxidize/oxidize-c test

Build the CUDA backend

$ make -C oxidize/oxidize-c cuda
Inspect oxidize-c on GitHub

Benchmarks you can rerun

Each backend runs back to back on the same GGUF and prompt. On our hardware, oxidize and miniforge beat stock llama.cpp.

Run the benchmark

Relative tokens per second, fastest run = 100

  • oxidize100
  • miniforge94
  • Ollama78
  • llama.cpp62

Ryzen 7 PRO, Qwen3.5-4B Q4, one prompt.

Side research

The Prion Atlas: Substrate, Detection, and Getting Someone to Order the Test

Nucleated-polymerization thresholds for prion disease, a three-lever therapeutic protocol evaluated with Monte Carlo simulation, and why diagnostic delay and RT-QuIC ordering decide outcomes.

Field
Medical preprint
Published
Sep 2026
DOI
10.5281/zenodo.22068068
License
CC BY 4.0
Read the preprint
llama.cppGGUFC11TritonConvexE2BHugging FacePyTorchMiniMaxOllama

Common questions

LLM performance. Quantization with turboquant, local inference stacks in oxidize, miniforge, and turboforge, and cross-backend benchmarks you can rerun. We also ship agent infrastructure like openagentmail.

Work with the lab on faster inference

Collaboration on quantization, local inference, or benchmark work starts with an email to jackson@email.zapdev.link.

ContactFund the research