Models & licenses
Smritsavant’s on-device intelligence is built on openly-licensed models. You meet them as capability tiers sized to your hardware — the app never surfaces model names, and neither does this page. What we disclose per tier is what actually matters for your machine and your rights: the parameter class, the quantization (download size), and the license. The exact model identities are recorded once, in the EULA’s third-party model schedule — the agreement’s fineprint.
Models are downloaded to your device from their official repositories on first use — they are third-party software, licensed by their creators under the licenses below, not by us. Our license covers the app; theirs cover the weights.
Chat tiers (current release)
Tier pins are per app release — a future version may upgrade a tier’s underlying model, and this page and the EULA schedule change with it (see the changelog).
Tiers are separated by memory class — how much model your machine can actually hold — and the ladder differs between Apple Silicon (unified memory) and dedicated-GPU PCs, because the physics differ.
Apple Silicon (unified memory)
| Tier | Memory | Parameters | Quantization | License |
|---|---|---|---|---|
| Lite | 16 GB | ≈4B | Q4_K_M (~2.7 GB) | Apache 2.0 |
| Standard | 24 GB | ≈9B | Q4_K_M (~5.7 GB) | Apache 2.0 |
| Moderate | 32 GB | ≈26B mixture-of-experts | MXFP4 (~16.6 GB) | Apache 2.0 |
| Supreme | 36 GB | ≈27B dense — quality opt-in; Moderate stays the default | Q4_K_M (~16.8 GB) | Apache 2.0 |
| Ultra | 48 GB+ | ≈35B mixture-of-experts (default) · ≈31B dense (quality option) | MXFP4 (~21.7 GB) · Q4_K_M (~18.3 GB) | Apache 2.0 |
8 GB Macs sit below the local-chat floor: voice, captions, and search still run fully on-device (see the always-on suite); chat uses your own cloud key or Apple Intelligence where supported.
Dedicated-GPU PCs (VRAM)
| Tier | VRAM | Parameters | Quantization | License |
|---|---|---|---|---|
| Lite | 8 GB | ≈4B (default) · ≈9B (quality option, see note) | Q4_K_M (~2.7 GB) · Q4_K_M (~5.7 GB) | Apache 2.0 |
| Standard | 12 GB | ≈9B | Q4_K_M (~5.7 GB) | Apache 2.0 |
| Moderate | 16 GB | ≈9B | Q6_K (~7.5 GB) | Apache 2.0 |
| Supreme | 24 GB | ≈26B mixture-of-experts | MXFP4 (~16.6 GB) | Apache 2.0 |
| Ultra | 32 GB | ≈35B mixture-of-experts (default) · ≈27B dense (quality option) | MXFP4 (~21.7 GB) · Q4_K_M (~16.8 GB) | Apache 2.0 |
A note on the 8 GB quality option: the ≈9B option needs nearly the whole card — about 7 GB even with the context trimmed and speech recognition kept entirely on the CPU — and a card that also drives your display gives up roughly 2 GB of that to the desktop, compositing, and driver buffers before a model loads. So in the current release the option appears where the display runs on the integrated GPU and the card is fully free for the model; every 8 GB card runs the ≈4B default at full speed regardless. This is memory arithmetic, not product segmentation, and the gate gets re-examined as we validate more real configurations.
PCs without a dedicated GPU are gated by memory bandwidth, not capacity — plenty of slow RAM still runs a big model at an unusable crawl — so the app doesn’t guess from a hardware list or a chipset name. A ~10-second speed check on first launch measures what your memory and integrated graphics actually deliver, and that measurement decides the rung: the ladder is open upward, not capped by category. The new class of high-bandwidth unified-memory PCs — APUs and integrated-graphics boards with fast soldered RAM, the kind now marketed for gaming and AI — can climb the same memory rungs as the tables above with the fast mixture-of-experts options; a typical dual-channel desktop or laptop lands at Lite; machines below the bar keep the full voice + caption + search experience with transcription queued rather than live, and chat via your own key. Dense quality options remain Apple Silicon / dedicated-GPU territory: on shared memory they’d generate at a speed we wouldn’t ship.
A few honest notes on the tables. On the GPU ladder, Standard and Moderate are deliberately the same brain at different numerical precision — Moderate buys cleaner long-form output, not a bigger model. On Apple Silicon, Moderate instead jumps to a mixture-of-experts model: it stores a ≈26B model’s knowledge but activates only a small slice per word, so it fits like a big model and answers at small-model speed — that’s also why the biggest tiers default to mixture-of-experts and keep dense models as opt-in quality choices. And a tier’s hardware gate is physics, not upselling: a ≈27B dense model’s working set simply doesn’t fit a 32 GB machine with room left for the system, so the app won’t offer it there — quality opt-ins appear exactly where they physically fit. An on-device model is capable, but it is not a frontier cloud model, and we won’t pretend otherwise. Frontier-class answers are what the bring-your-own-key cloud option is for.
On Apple devices that support Apple Intelligence, chat can also run on Apple’s foundation models — system-managed, nothing downloaded by us. On Android, chat uses your own cloud key.
The always-on suite (every install)
Voice, captions, and search aren’t tiered — every desktop machine gets them, including the ones below the chat floor. What varies is only engine size and whether transcription is live or queued:
| Role | Size | License |
|---|---|---|
| Semantic memory & search (a multilingual sentence embedder) | ~22 MB | Apache 2.0 |
| Voice — language identification (a tiny speech model that routes each recording to the right engine) | ~150 MB | MIT |
| Voice — speech-to-text engine (multilingual; bigger machines get the larger one) | ~0.8 GB or ~2.2 GB | Apache 2.0 |
| Voice — Tamil, Telugu, Kannada, and Malayalam speech packs (specialised fine-tunes; each downloads on first use of its language) | ~540 MB each | Apache 2.0, attribution required |
| Voice — fallback transcription (a compact multilingual speech model) | ~0.5 GB | MIT |
| Image captions (a compact vision model, batched in idle moments) | ~1.7 GB | Apache 2.0 |
Language coverage never shrinks with hardware: the same identification-and-routing suite runs on every machine — the speech engines are CPU-capable — and on machines below the speed bar the app records now and transcribes in the background, updating the note when the transcript is ready, instead of dropping languages to stay live.
What the licenses mean for you
- Apache 2.0 / MIT are permissive open licenses: free to use, including commercially, on your own machine. The full attribution — which model, from whom, under which text — is preserved in the EULA’s third-party model schedule and in-app (Settings → About → Third-party licenses).
- The Tamil, Telugu, and Kannada speech packs were fine-tuned by the Speech Lab at IIT Madras (work supported by India’s Bhashini programme); the Malayalam pack by an independent open-source contributor. Both are Apache 2.0 with attribution, which the EULA schedule and in-app credits carry in full.
- If you point the app at your own models (via your Ollama server) or a cloud provider (your API key), those run under whatever terms you’ve accepted with that provider — outside this page’s scope.
- We deliberately pin only models whose licenses permit this distribution model. If a future tier ever uses a model with additional use restrictions, this page will say so explicitly, before you download it.