The Quantization Challenge
Take an open-weight language model, make it run in 8 GB of memory at a 4,096-token context, and keep as much of its quality as you can. Any method counts: GGUF, GPTQ, AWQ, mixed precision, pruning plus quantization, your own idea.
Season 1 timeline
Registration open now
Register to receive the starter kit: baseline scripts, the evaluation harness and the public validation prompts.
1 December 2026 — submissions open
Prize pool and final prize split published. Leaderboard goes live with the public validation score.
28 February 2027, 23:59 UTC — submissions close
Each team selects up to two final submissions.
March 2027 — results
Final scores on the hidden test set, verification of winners and publication of the top write-ups.
How entries are scored
Every submission must pass a memory gate, then competes on quality.
| Stage | Rule |
|---|---|
| Memory gate | Peak memory, measured by the organizers' harness, at most 8.0 GiB with a 4,096-token context and batch size 1. |
| Quality score (100%) | Average of task scores on the hidden test set, each divided by the same base model's BF16 score. 100 means no measurable loss. |
| Tie-breakers | Lower peak memory, then earlier submission time. |
| Best write-up award | Judged separately on clarity, reproducibility and insight (1–5 scale per criterion). |
Allowed base models are listed in the starter kit and limited to openly licensed models whose licence permits modification and redistribution.
Planned prize structure
The pool is split 50% first place, 25% second, 15% third and 10% best write-up. Winners also get a featured case study on Quantization.com. One cash prize per team.
Register your team
Official rules (summary)
- No purchase necessary. Entry is free. Void where prohibited or restricted by law.
- Eligibility: individuals at least 18 years old or the age of majority in their place of residence. Organizers, judges, sponsors' staff working on the contest and their immediate families are not eligible for prizes. Residents of countries subject to comprehensive sanctions cannot receive prizes.
- Original work: submissions must be created during the contest period by the team. Public open-source tools are allowed and must be credited. Disclose any AI tools used.
- Reproducibility: include a public repository under an OSI-approved licence (MIT or Apache-2.0 recommended) with a README that reproduces the result.
- Fair play: no training, tuning or calibration on the hidden test set; no manipulation of the measuring harness. One account per person.
- Verification: potential winners must respond within 7 days and provide proof of identity and eligibility, and any tax forms required to pay prizes. Unclaimed prizes may go to the next eligible team.
- Ownership: you keep all rights to your work. You grant Quantization.com a non-exclusive, royalty-free licence to describe, display and link to your submission and write-up.
- Changes: the organizers may amend these rules or cancel the season if fairness or legality requires, with notice on this page. Decisions of the judges are final.
Full rules, privacy and legal terms
Organizer. Quantization.com. Contact through the contact form.
Contest type. A skill-based competition; chance plays no role in selecting winners. Where local law requires registration, filings or a skill-testing question (for example for residents of Québec or elsewhere in Canada), the organizer will meet those requirements before prizes are awarded.
Prizes. Cash prizes are paid by bank transfer or another method agreed with the winner, in USD, within 60 days of verification. Winners are responsible for taxes on their prizes. Prizes are not transferable.
Publicity. Winners agree that their team name, country and submission summary may be published. Personal contact details are never published.
Privacy. Registration data is used only to run the contest and contact participants, as described in the privacy policy.
Liability. To the extent permitted by law, the organizer is not responsible for lost, late or corrupted submissions, or for technical failures outside its control.
Governing law. These rules are governed by the laws applicable at the organizer's place of business, without prejudice to mandatory consumer protections where you live.
Sponsor a prize
Put your company in front of the engineers who care most about efficient inference. Title, category and compute-credit sponsorships are available, each clearly credited on the challenge page, leaderboard and winner write-ups. Sponsors never influence scoring.
Also interested in the whole site or domain? Contact the owner.
Propose a future challenge
Ideas we are considering: an 8-bit KV-cache accuracy race, a "smallest model that passes" coding challenge, an ADC dither listening test and a music-quantize groove contest. Pitch yours.