# sqliteai/warp

Run the full 2.78-trillion-parameter Kimi K3 model beyond available RAM by streaming activated weights directly from NVMe. A dependency-free, embeddable C inference engine.

Repository: https://github.com/sqliteai/warp
Canonical: https://ross.abutalabs.com/products/sqliteai-warp
Homepage: https://sqlite.ai
Language: C
License: Apache-2.0
License Family: permissive
Last push: 2026-08-24T09:16:44+00:00

## Health v2 (maintenance only)
Score: 80/100 (v2, computed 2026-09-03T02:20:16.233290+00:00)
- activity 99, release rhythm 99, longevity 2
- inputs: {"age_days": 36, "days_push": 9, "days_rel": 9, "gap_med": 1, "n_releases_24m": 10}
- flags: young
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 2284, forks 168 (observed 2026-08-28T04:06:33.903522+00:00)

## What it is
WARP is a dependency-free, embeddable C inference engine that runs trillion-parameter mixture-of-experts models like Kimi K3 by streaming activated expert weights directly from NVMe storage instead of holding them in RAM. It keeps shared weights in memory, uses a bounded expert cache, and achieves full-model local inference on consumer hardware such as a 64 GB MacBook Pro.

## Use cases
- run the full Kimi K3 model locally on a 64 GB machine
- stream LLM weights from NVMe when the model exceeds available RAM
- embed a frontier-model inference engine in a C application with no dependencies
- experiment with mixture-of-experts weight paging and expert caching
- run large MoE models on consumer hardware without distillation or pruning

## When to choose
- you want to run a huge MoE model locally and it does not fit in RAM
- you need a dependency-free, embeddable C inference engine
- you have fast NVMe storage and enough RAM for the shared trunk plus a cache
- you want to experiment with weight-streaming inference techniques

## When to avoid
- you need high token throughput for interactive chat (K3 decodes at ~0.6 tok/s)
- you want broad model support or a mature ecosystem like llama.cpp
- your machine has less than 64 GB RAM for the largest models
- you need GPU-accelerated inference or multi-GPU serving

## Facets
- artifact type: library
- maturity: active
- function: llm-inference, machine-learning
- domain: large-language-models, machine-learning, developer-tools
- platform: cross-platform, cli
- tags: mixture-of-experts, nvme-streaming, quantization, embeddable, c, local-inference, kimi-k3, macos, linux

## Member repositories
- sqliteai/warp (main) score 80

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:06:33.903522+00:00.
- Health v2: computed from the inputs above; adoption is never an input.
- Inferred fields (summary, facets, guidance): AI-extracted, prompt v1, taxonomy v1, on 2026-08-30T02:41:06.810480+00:00, confidence not recorded.
  - readme: https://github.com/sqliteai/warp (fetched 2026-08-28T04:06:33.903522+00:00, sha 0f4962e58884)
  - homepage: https://sqlite.ai (fetched 2026-08-29T10:21:26.920475+00:00, sha 99d8968ff13d)
  - site_page: https://www.sqlite.ai/pricing (fetched 2026-08-29T10:21:26.929772+00:00, sha f02c43a2a836)
  - site_page: https://www.sqlite.ai/changelog (fetched 2026-08-29T10:21:26.932100+00:00, sha 30a18014da99)
- Data as of 2026-08-30T08:39:29.467469+00:00.