Alphabell.

Home / Mini-grants

Mini-grants

Small, low-paperwork awards for the kind of AI work that doesn't get easier at scale — sharp hypotheses, clean replications, focused experiments that need weeks of thinking more than they need more GPUs. $500 to $25,000. Single short application. Decisions in weeks rather than months. Inspired by the Emergent Ventures model.

Cohort
7 — open through 30 Jun 2026
Range
$500 – $25,000
Decision
14-day median

What we fund

  • One-off research projects — a paper, a write-up, a short focused experiment
  • Open-source tooling for the open AI stack
  • Reproductions and audits of published work
  • Conference travel for first-time presenters without institutional support
  • Paid contributor time on a member-led project
  • Reading groups, local workshops, mentor honoraria
  • Datasets, paid annotation hours, or small equipment buys

What we don't fund

  • Salaries for full-time employees of frontier AI labs
  • Pure compute purchases — we already have compute, use ours
  • Closed-source research with no public output
  • Tuition for accredited degree programs

How decisions are made

Each application goes to two reviewers, drawn at random from the L7+ pool. We don't ask for a CV. We ask: what's the smallest thing that would change your mind about this idea, and how much money would it cost to find that out?

Reviewers are paid per accepted review, not per decision. Median time to a decision: 14 days. We tell you why we said no.

Cohort 7 Open

Apply through 30 Jun 2026

Submit any time. Reviewers work the queue continuously.

$500 – $25,000· 14-day median
Cohort 8 Coming up

Opens 1 Jul 2026

Opens automatically when Cohort 7 closes. Same terms, same process.

$500 – $25,000· 14-day median

Start a mini-grant application

Public record

Every grant we've made

We publish every disbursement: name, country, amount, month, and a one-paragraph description of what the money paid for. 16 entries below; full archive on GitHub.

  • Aditi RaoBengaluru, IndiaMay 2026$5,000

    An open evaluation harness for tool-use under partial observability. Drop-in replacement for hand-coded react/plan benchmarks — already used by four cohort-6 teams.

  • Tomasz WłodarczykKraków, PolandMay 2026$3,200

    Reproducible benchmark for KV-cache compression across six long-context open models. Published under MIT; the SGLang authors cited it within a month.

  • Mariana FurtadoPorto, PortugalApr 2026$8,500

    Hosts a weekly Lisbon–Porto interpretability reading group. Grant covers room rental, food, and Eleuther-style mentor honoraria for nine months.

  • Cheung Wai-LinHong KongApr 2026$12,000

    Adversarial-prompt corpus for Cantonese tool-use — the first public dataset of its kind. 18k labelled examples, dual-released to Hugging Face and OpenReview.

  • Dineth KarunaratneColombo, Sri LankaApr 2026$2,400

    Independent reproduction + bug-bounty on the Apollo router-distillation paper. Found three numerical bugs the original authors had missed.

  • Priya AnandToronto, CanadaMar 2026$4,000

    Multimodal eval set for blind-spot detection in vision-language agents. 600 hand-curated trick questions, reproducible scoring pipeline.

  • Yusuf Demirİstanbul, TürkiyeMar 2026$1,500

    Travel + lodging to present at a NeurIPS workshop on agentic RL. Yusuf is 19 and unaffiliated with any university.

  • Sofia KarvelasAthens, GreeceMar 2026$7,200

    Mechanistic study of refusal directions in small open-weight models. Pre-registered and published with full notebooks.

  • Vihaan MehtaMumbai, IndiaMar 2026$3,800

    Tooling to translate Hugging Face datasets into Indic-language evaluations. Now wired into 11 community fine-tunes.

  • Eleanor PakSeoul, South KoreaFeb 2026$5,500

    Reproducible benchmark for sparse-autoencoder dictionary quality on small LMs. Used in two follow-up papers already.

  • Henrik LindqvistGöteborg, SwedenFeb 2026$2,000

    Conference travel + lodging to present a poster on residual-stream causal scrubbing at ICLR.

  • Anya PetrenkoLviv, UkraineFeb 2026$4,800

    Twelve weeks of full-time work on a survey of post-training routing techniques. Draft now in OpenReview.

  • Kenji MaruyamaKyoto, JapanJan 2026$6,000

    Hand-rolled implementation of the PIQA + ARC eval set in JAX for low-resource teams; ships with verified reference numbers.

  • Adaeze OkekeLagos, NigeriaJan 2026$3,500

    Public dataset of African-language code-switching prompts; first of its kind for tool-use. Released under CC-BY.

  • Léa BouchardMontréal, CanadaJan 2026$5,800

    Single-author paper on attention-sink dynamics in long-context models, with code + checkpoints released alongside.

  • Daniel BergerBerlin, GermanyDec 2025$9,000

    Author of `agentcheck`, a lightweight static analyser for agent-tool schemas — now an alphabell standard for new submissions.