Your text never leaves your device
Snippets, prompts, and generated output stay local because inference runs through Ollama on your own machine. No third-party server ever sees a keystroke or an expansion.
Local AI & Benchmark, Measured Honestly
On-device AI features and the independent data behind them. Every model spec on this page comes from the official Ollama library, every app fact is measured per release, and nothing is invented. No cloud API, no telemetry, and no per-token fees.
AI FEATURES
Built-In Intelligence
A fine-tuned model hub and a local Ollama engine make AI practical, private, and fast.
Smart Shorthand
Expand abbreviations into natural, complete sentences using a fine-tuned Qwen 1.5B model, purpose-built for shorthand and running entirely on-device.
AI Snippet Creator
Describe what you need and let AI generate the snippet for you. It lands straight in your library, with no cloud round-trip.
Semantic OmniSearch
Find snippets by meaning, not just keywords. Local vector embeddings via nomic-embed-text rank results by cosine similarity.
One-Click Model Hub
Discover and download embedding, text generation, and shorthand models in one click. Snipset handles Ollama engine setup and lifecycle automatically.
HOW IT WORKS
Your Machine Is the AI Server
Snipset runs a local Ollama instance and connects automatically on port 11434. Text generation, embeddings, and semantic ranking all execute on your hardware. There is no cloud API, no account required for AI, and no telemetry. Your proprietary code and private data stay where they belong.
- • CPU is enough. No GPU required; a discrete GPU simply makes generation faster.
- • Zero telemetry. Prompts and generated text never leave your machine.
- • Automatic lifecycle. Ollama setup, model downloads, and legacy model cleanup are handled for you.
BENCHMARK DATA
Sourced, Measured, Reproducible
Every figure on this page is either sourced from the official Ollama model library or produced by a release pipeline. Nothing is copied from a vendor press release, and nothing is invented.
COMPETITIVE ADVANTAGES
Why local AI is a competitive advantage
Against cloud AI assistants and legacy snippet tools, running inference on your own machine changes latency, privacy, cost, and the footprint of the product. These are the four structural advantages, and each one can be verified.
One-time license, no metered AI fees
The AI runs on hardware you already own. Snipset is a one-time license, not a per-seat SaaS plan with an added per-token cloud bill.
A footprint you can measure
Installer size and memory are published per release from a reproducible pipeline on the /performance page, so the footprint is evidence, not a claim on a spec sheet.
Any model you want, fully offline
Choose a 1.5B shorthand model or an 8B code model from the catalog below, then disconnect. It keeps working, because nothing depends on a cloud endpoint.
MODEL CATALOG
On-device models, with their official specs
Sizes, context windows, and parameter counts are the official Ollama library values. Throughput is only shown once a reproducible run is published, and until then it is marked pending.
Alibaba
Qwen 2.5 (1.5B)
ollama pull qwen2.5:1.5bOfficial model size986 MB
Context window33K
Parameters1.5B
Measured throughputPending release run
Official Ollama library →Microsoft
Phi-3 Mini (3.8B)
ollama pull phi3:miniOfficial model size2.2 GB
Context window131K
Parameters3.8B
Measured throughputPending release run
Official Ollama library →Hugging Face
SmolLM2 (1.7B)
ollama pull smollm2:1.7bOfficial model size1.8 GB
Context window8K
Parameters1.7B
Measured throughputPending release run
Official Ollama library →Cloud option, not on-device:
gemma4:31b-cloud streams from Ollama Cloud, so no model file is downloaded locally and your prompts leave the machine. On-device Gemma 4 variants such as gemma4:31b (20 GB) keep every token local and require a high-VRAM workstation.HARDWARE GUIDANCE
Match the tier to your machine
Recommended ranges follow directly from the official model sizes above. Exact throughput still depends on your hardware and will be measured per release.
Tier 1: CPU-only
Standard laptops and PCs
- ✓ Shorthand and embedding models up to 1.8 GB
- ✓ 8 to 16 GB system RAM
- ✓ Zero GPU configuration required
Tier 2: Entry GPU / Apple Silicon
Modern creator laptops
- ✓ Models up to 3.8 GB, smooth long-form generation
- ✓ 4 to 6 GB VRAM or an M-series chip
- ✓ Good fit for the 7B tier on capable machines
Tier 3: Dedicated GPU
Workstations and high-end
- ✓ Full 8B code generation tier (4.9 GB model)
- ✓ 8 to 16 GB VRAM, 32 GB system RAM
- ✓ Large-batch semantic embeddings
RELEASE MEASUREMENTS
Measured on every release
These are Snipset app facts generated by the reproducible pipeline documented on the /performance page. Unmeasured values stay marked pending; they are never filled with a guess.
Installer MSI
12.4 MB
measured 2026-09-06
Setup EXE
9.6 MB
measured 2026-09-06
Tray-idle memory
Pending release runsampled on the release runner
Engine match p50
Pending release runmicro-benchmark, per release
Full methodology and the latest measured engine and process numbers:Snipset /performance →
COMPETITIVE ADVANTAGE
Snipset vs Cloud AI vs Legacy Tools
A compact, qualitative comparison. Every row reflects Snipset's actual architecture, and the detailed head-to-head table lives on the /compare page.
| Dimension | Snipset (local AI) | Cloud AI assistants | Legacy snippet tools |
|---|---|---|---|
| Architecture and data | |||
| Where the AI runs | On your device, via Ollama | Vendor servers | No AI, or a paid cloud add-on |
| Where your text goes | Stays on your device | Uploaded for inference | Depends on sync settings |
| Cost and independence | |||
| Pricing | One-time license | Per-token or per-seat | Per-seat subscription |
| Offline use | Fully offline | Requires internet | Mostly offline |
| Runtime footprint | Native, measured per release | Browser or heavy app | Often Electron apps |
| Model choice | 1.5B to 8B, your pick | Fixed by the vendor | None |
For the full competitor table across memory, startup time, and pricing:How Snipset compares →
METHODOLOGY
Where these numbers come from
The rule for this page is simple: sourced or measured, never invented.
- Model specs are the official values on the Ollama model library, retrieved2026-09-05. Each card links to its tag page so the size, context window, and parameter count can be verified in one click.
- App facts (installer size, and eventually memory and engine timings) are generated by the same pipeline that releases the product, documented on the /performancepage, and refreshed on every release.
- Token throughput and latency are not published until a reproducible run exists on reference hardware. Until that run happens, the card shows Pending release run instead of a number, because a fake number would be worse than none.
FREQUENTLY ASKED QUESTIONS
Honest answers about benchmarks
The questions an engineer, or an investor, asks before trusting a benchmark.
Why are there no tok/s numbers on this page yet?
Because none have been measured honestly. Token throughput and first-token latency depend on specific hardware, and this page refuses to publish invented numbers. The moment a reproducible run exists on reference hardware, the results appear here with the date, the command, and the machine that produced them.
Where do the model sizes and context windows come from?
Directly from the official Ollama model library (ollama.com/library), which publishes the quantized download size, context window, and parameter count for every tag. This page records the retrieval date and links each card to its official tag page so anyone can verify the figures.
Do I need a GPU to use local AI in Snipset?
No. The 1.5B to 3.8B shorthand models and the embedding models run comfortably on CPU-only machines with 8 to 16 GB of RAM. A 4 to 6 GB GPU extends you to mid-size generation, and 8 GB or more VRAM unlocks the 7B to 8B tier. The hardware section above maps each tier to the official model size.
Which model does Snipset ship by default?
Snipset ships a fine-tuned shorthand model built from the Qwen 2.5 family and served as a GGUF file, with a catalog of optional Ollama models like the ones profiled here. You can add or swap any compatible model without changing your snippets.
Do model files or my prompts ever leave my device?
No. Inference runs locally through Ollama on your own machine. Your snippets, prompts, and the generated text never go to a third-party server, which is the core architectural advantage the competitive matrix above summarizes.
How often are these figures refreshed?
The Ollama library specs are re-verified whenever a model card is updated, and the retrieval date on the trust bar tracks that check. The Snipset app facts in the measured band are regenerated on every release by the same pipeline documented on the /performance page.