The Janas repository publishes GPL-3.0 code, SSD expert streaming and author-run benchmarks.
Janas streams model experts from SSD on memory-limited computers
2026-09-20Frontier watchIGNORE
Janas is a GPL-3.0 runtime that streams mixture-of-experts weights from SSD so models larger than system memory can run locally. The repository claims large speedups over llama.cpp across 13 quantized models, but those comparisons are author-run and the project has no independent quality, stability or hardware-normalized evaluation.
What this means for you
SSD expert streaming is useful when memory is scarce, but author-only speed claims and a young runtime do not yet justify integration work. Treat community attention as a discovery signal and wait for reproducible, hardware-matched comparisons.
FRONTIER WATCH
Public artifacts and validated results are different.
The recorded fact is the public work and its artifacts. Author-reported outcomes do not establish independent reproduction. Next checks include usable code, task conditions and third-party validation.
Evidence & related material
Technical details & scores
| License | GPL-3.0 |
|---|---|
| Approach | stream MoE experts from SSD |
| Author Claim | faster than llama.cpp on 13 Q4_K_M models |
| Adoption Snapshot | 286 GitHub stars and 18 forks at review time; variable community signal |
| Limits | no independent quality, stability or hardware-normalized evaluation |
| Time Basis | repository created 2026-09-20T09:55:09Z |
Technical impact 2
Engineering value 2
Adoption signal 2
Maturity 1
Verification cost 4
Risk 3