ds4

What it does

An editorial selection.

"DwarfStar", a small native inference engine from Redis creator antirez for DeepSeek V4 Flash/PRO, GLM 5.x and Qwen3.8 on Metal, CUDA and ROCm.

Why it is listed

Focused, self-contained C code tuned to run a few excellent models well on hardware people actually own.

GitHub describes it as: DeepSeek 4 Flash and PRO local inference engine for Metal, CUDA and ROCm

Counts were observed on 27/10/2026, 23:15:40 UTC. Observations older than 36 hours are excluded from current movement rankings.

Repository at a glance

Stars: 24,116
Forks: 2,339
Open issues: 306
Merged PRs: 38
Commits, last 30 days: 31
Stars, 7 days: unavailable
Stars, 30 days: unavailable

Counts observed on 27/10/2026, 23:15:40 UTC. Growth uses complete dated windows; see Methodology.

History

History: 2 observations, 2026-10-27 to 2026-10-28. Dots are observations; lines connect consecutive daily observations only.

Project details

Language: C
Licence: MIT
Created: 6 May 2026
Last push: 23 Oct 2026
Latest release: none published
Category: Local AI Inference
Topics: none listed
Observed: 2026-10-27 23:15:40 UTC

Explore and get started

Use the project's own documentation for supported platforms, dependencies and installation steps. The README and release notes are maintained by antirez.

Read the README and setup instructionsFeatures, prerequisites, examples and documentation links.
Browse releases and changelogsPublished versions and the changes made between releases.
Check licence and usage termsMIT is the licence reported by GitHub; read the repository terms before use.
Read pull requestsProposed changes and development discussions.
Browse reported issuesKnown problems, feature requests and maintainer responses.

Development context

These are observations of work in the repository. They do not establish quality, security or support commitments.

Total commits on default branch: 829
Open pull requests: 539
Closed issues, lifetime: 146
Median time to close an issue: 4.4 days
Median time to merge a pull request: 2.2 days

Close and merge times use up to 20 most recently updated closed issues or merged pull requests. They omit unresolved work and are context, not a maintenance score. Sample limits.

Other projects in Local AI Inference, ordered by shared topics and then name. These are discovery links, not equivalence claims.