AI Potluck
Contribute
Back to Gap Map Model components / Fine-tuning code

Lamini

AMD
closed / Overall score: n/a

Lamini is an enterprise fine-tuning and inference platform built around Mixture of Memory Experts (MoME), a technique that trains many LoRA-style memory experts per model to target hallucinations. It ships as an on-demand or reserved hosted service, or as a self-managed deployment on a customer's own GPUs, with a Python client library as the entry point to the tuning engine in every tier.

Reported as an AMD acquisition in 2025, but the sources show only a leadership move: co-founder Sharon Zhou and several colleagues joined AMD, while Lamini's own documentation still lists its service tiers as live with no wind-down language.

Openness

1 high confidence
1.0
license
Proprietary
source
closed
engine
closed(hosted memory-tuning/fine-tuning service on Lamini/AMD servers)
client-sdk
open(Apache-2.0, lamini-ai/lamini)
data
closed
training-code
closed
access
API-key-gated

Proprietary managed fine-tuning platform; no source is published for the engine in any of the on-demand, reserved, or self-managed deployment tiers, and the only open artifact is a thin API client SDK, so the product itself remains closed.

  • https://docs.lamini.ai/about/ recorded 2026-09-22

    States "Lamini Platform is available in three different deployment models": On-Demand (fully-managed training/inference at app.lamini.ai, pay-as-you-go), Reserved (dedicated GPUs on Lamini's own infrastructure, per-GPU pricing), and Self-Managed (run in the customer's own environment/GPUs -- on premise, VPC, or air-gapped -- per-GPU pricing). Every tier is a paid deployment of Lamini's own platform; none exposes a published repository for the engine.

  • https://docs.lamini.ai/LICENSE/ recorded 2026-09-22

    Lists only third-party open-source notices bundled into the self-managed distribution -- vLLM (Apache-2.0), LlamaFactory (Apache-2.0, credited for supporting "MoME training"), litellm (MIT), and Ray (Apache-2.0 modifications) -- with no license grant for Lamini's own platform code anywhere on the page.

  • https://github.com/lamini-ai/lamini recorded 2026-09-22

    Repo title/description read "The Official Python Client for Lamini's API"; GitHub's own repo metadata records it public, not archived (isArchived:false), licensed Apache-2.0, with website field pointing to https://lamini.ai/. No fine-tuning or inference engine code lives in the repo -- only the API client.

Adoption

not assessed

Lamini publishes no usage, customer or download figure for its platform, and the founders' claim of large-scale production deployment describes their earlier careers rather than this product. The only public artifact is a thin API client, so adoption is left unscored rather than inferred from press coverage.

  • https://docs.lamini.ai/about/ recorded 2026-09-28

    The About page says the founders "shipped LLMs in production to over 1 billion users" as a claim about their prior careers, and gives no download, customer or active-user count for the Lamini platform; searched the body for download, customers, users served and active users tied to the product: none.

Capability

not assessed

Lamini markets memory-tuning with claims of about 95% factual accuracy and 32x compression, but the arXiv paper behind those figures was withdrawn by its own authors, and the closed engine cannot be independently checked. No score is assigned rather than one built on an unverifiable claim.

  • https://arxiv.org/abs/2406.17642 recorded 2026-09-28

    The abstract page for the Lamini memory-tuning paper carries the notice "This paper has been withdrawn by Gregory Diamos" and a submission history entry ending "(withdrawn)".

  • https://docs.lamini.ai/about/ recorded 2026-09-28

    The About page lists "95% accuracy on factual tasks" and "32x model compression" as vendor claims for memory tuning, with no benchmark harness or reproduction linked.

  • https://raw.githubusercontent.com/lamini-ai/lamini/main/README.md recorded 2026-09-28

    The repository README says "This python package is a client + SDK compatible with the Lamini API", so only the API client is public; searched for engine, training code and benchmark harness: none.

Verified 2026-09-22