Supported
Executed by AI Stays Local on our own hardware and accepted. Evidence is published on the benchmarks page.
Models
What we have actually run
A model is supported when AI Stays Local has executed it and accepted the result. An adapter existing in an upstream engine is not support; it is an adapter. Every entry states total and active parameters separately, because they are different numbers.
In validation
Being brought up and measured now. It is not a capability of the product, and it may not survive validation.
Planned
Intended for a future release. Nothing has been run yet.
Upstream reference
Demonstrated by an upstream project, not reproduced by AI Stays Local. Evidence that the physics works, not evidence that we can do it.
OLMoE 7B-A1B
Supported7B total1B active per tokenMixture-of-Experts
- Hardware class
- lightweight
- Quantisation
- Recorded with the benchmark
- Runtime
- AI Stays Local runtime
- Licence
- Apache-2.0 (model weights, per the publisher)
- Provenance
- Executed by AI Stays Local on NVIDIA DGX Spark, Linux ARM64, CPU inference.
- Evidence
- Benchmark record
LimitationsA small model. Strong on everyday drafting, summarising and questions about your own documents; it will lose to a large hosted model on hard reasoning.
Qwen3.5-122B-A10B
In validation122B total10B active per tokenMixture-of-Experts
- Hardware class
- professional
- Quantisation
- To be fixed during validation
- Runtime
- AI Stays Local runtime, multi-engine
- Licence
- Published by the model publisher; licence verified against the publisher repository before any result is published
- Provenance
- A published open-weight model from its publisher. AI Stays Local has not completed its own execution and acceptance, so nothing here is attributed to us.
- Evidence
- No published record yet
LimitationsUnder validation on GPU. It is not operational, we do not claim it, and no third-party benchmark for it is published here as ours.
GLM-class model, approaching 744B total parameters
Upstream reference~744B totalA small active fraction per token; other experts streamed from local storageMixture-of-Experts with expert streaming
- Hardware class
- extreme
- Quantisation
- As reported upstream
- Runtime
- Upstream Colibrì project
- Licence
- Per model publisher
- Provenance
- Reported by the upstream Colibrì project. Not reproduced by AI Stays Local, on any hardware, at any time.
- Evidence
- Benchmark record
LimitationsUpstream research evidence that expert streaming scales. It needs a great deal of local storage, reported generation is far slower than interactive cloud chat, and it is not a capability of this product.
The upstream entry is included because it shows the approach is physically possible, not because it is something this product does. It has not been reproduced by AI Stays Local on any hardware.