The Austin Python Meetup

New location at Brazos Place @ 800 Brazos st. us the entrance on brazos street not 8th.
This is a Hybrid event both in-person and virtual. The in-person event is hosted at Antler downtown Austin.

Address:
Brazos Place @ 800 Brazos st,
The entrance is off of Brazos st.

Talk title: Manu Nicholas -Your Quantization Format Is Not Free: Measuring What On-Device LLMs Actually Cost

Talk description: When you download a quantized model you pick it by its filename. Q4_K_M, Q3_K_M, IQ4_XS. That name reads like a specification, but it is closer to a request. I built eight GGUF files from a single FP16 source and looked inside each one, and only three contained the format their name claims. Q3_K_M contained no Q3_K blocks at all. Q4_K_M at 0.5B parameters turned out to be 12 percent Q4_K by tensor count, storing 5.52 bits per weight against a nominal 4.5.
The gap is not academic. Two files carrying the same label decode up to 38 percent apart on the same machine, and the effect is large on a weak ARM core while nearly vanishing on x86, which is why it goes unnoticed on a developer workstation and then surprises people on a Raspberry Pi or a phone. This talk is about measuring that yourself. I will walk through the Python tooling I use to inspect a model file, show the numbers across three microarchitectures, and cover which formats sit off the efficiency envelope on every core I have tested. Everything is open and pip-installable, and reproducible on hardware you already own. Bring a model you actually use and we can look at what is really inside it.

Bio: Manu Nicholas Jacob is a hardware engineer at Dell in Austin, where he works in root-cause engineering on enterprise AI servers, looking at why systems fail and what the data actually says about it. He studied electrical and computer engineering at UMass Amherst.
Outside work he runs reproducible measurement studies on edge AI hardware, mostly Raspberry Pi 5 boards and laptops, on what really governs latency and energy: memory bandwidth, thermal limits, and quantization format. He maintains two open-source Python packages, llama-roofline and ml-systems-lab, reviews for the Journal of Open Source Software, and serves on the USENIX ATC 2026 artifact evaluation committee. He has a habit of publishing results that contradict his own earlier ones.
Happy to adjust length or depth for the room. Let me know what slot you have in mind and I will build to it.

Antler
Antler works with founders through our signature founder residency, adding value immediately, and taking no equity up front. We receive over 160,000 applications each year to build companies with us, and have grown to 18 funds with residencies in 30 cities around the globe. Founders also have the opportunity to pitch our Investment Committee to receive an up to $500K investment, allowing them to build for 12-18 months rather than attempting to raise before they have real traction.