// model_catalog
12 models that run on Mortise
Curated the hard way: we ran 29 candidates through the same on-device tests and kept the 12 that earned a place.
Every one is quantized for Apple Silicon and runs fully on-device. Grouped by where it runs, then by family — memory figures are on-device peaks, weights are download size.
// how_this_was_measured
Measured in Mortise against 72 deterministic tasks spanning factual recall, arithmetic, code reading, logic, instruction-following and extraction. Models run as 4-bit MLX conversions (3-bit for some), under this app's prompts, sampling and memory limits.
- Strongest
- Handled the widest range of the suite's tasks in this app.
- Strong
- Dependable across most task types, with occasional misses.
- Capable
- Solid on everyday tasks; loses ground on the harder ones.
- Limited
- Struggled with much of the suite under this app's prompts and limits.
- Not yet measured
- Not run deeply enough for us to say anything. It is an absence of data, not a low standing.
This describes how each conversion behaves inside Mortise. It is not a general benchmark of the underlying model, and quantized conversions do not represent the full-precision weights their makers published.
iPhone & iPad
8 modelsModels that fit within iOS's ~5 GB per-app memory budget. “Tight” models still run, but sit close to the limit on some devices.
Qwen 4
Qwen2.5 0.5B Instruct (4-bit)
Ultra-light chat and quick PDF Q&A; long answers cut off
- RAM peak
- 1.5–2.3 GB
- Weights
- 0.27 GB
- Context
- 32,768 tok
- Min Mac RAM
- 8 GB
Qwen3 0.6B (4-bit)
Fast reasoning, PDF Q&A, and doc lookup at a tiny size
- RAM peak
- 1.5–2.3 GB
- Weights
- 0.33 GB
- Context
- 40,960 tok
- Min Mac RAM
- 8 GB
Qwen3 1.7B (4-bit)
Reasoning plus accurate doc lookup in a light model
- RAM peak
- 2.1–2.9 GB
- Weights
- 0.92 GB
- Context
- 40,960 tok
- Min Mac RAM
- 8 GB
Qwen3 4B Instruct 2507 (4-bit)
Top all-rounder — reasoning, web search, PDF tools; passed every test
- RAM peak
- 3.3–4.1 GB
- Weights
- 2.12 GB
- Context
- 262,144 tok
- Min Mac RAM
- 8 GB
LFM (Liquid AI) 1
LFM2.5 1.2B Instruct (4-bit)
Fast, dependable chat and PDF/doc Q&A (Liquid AI)
- RAM peak
- 1.8–2.6 GB
- Weights
- 0.62 GB
- Context
- 128,000 tok
- Min Mac RAM
- 8 GB
SmolLM 1
SmolLM3 3B (4-bit)
Reasoning, PDFs, and doc lookup — passed every on-device test
- RAM peak
- 2.9–3.7 GB
- Weights
- 1.67 GB
- Context
- 65,536 tok
- Min Mac RAM
- 8 GB
Nemotron 1
Nemotron 3 Nano 4B (4-bit)
Reasoning, web search, and doc Q&A — near-perfect in tests
- RAM peak
- 3.2–4.0 GB
- Weights
- 2.24 GB
- Context
- 262,144 tok
- Min Mac RAM
- 8 GB
Ministral 1
Ministral 3 3B Instruct (4-bit)
Compact instruction-following chat and doc Q&A
- RAM peak
- 3.8–4.6 GB
- Weights
- 2.59 GB
- Context
- 262,144 tok
- Min Mac RAM
- 8 GB
Mac
12 modelsEvery catalog model runs on a Mac with enough memory — the card shows the minimum RAM each one needs.
Qwen 5
Qwen2.5 0.5B Instruct (4-bit)
Ultra-light chat and quick PDF Q&A; long answers cut off
- RAM peak
- 1.5–2.3 GB
- Weights
- 0.27 GB
- Context
- 32,768 tok
- Min Mac RAM
- 8 GB
Qwen3 0.6B (4-bit)
Fast reasoning, PDF Q&A, and doc lookup at a tiny size
- RAM peak
- 1.5–2.3 GB
- Weights
- 0.33 GB
- Context
- 40,960 tok
- Min Mac RAM
- 8 GB
Qwen3 1.7B (4-bit)
Reasoning plus accurate doc lookup in a light model
- RAM peak
- 2.1–2.9 GB
- Weights
- 0.92 GB
- Context
- 40,960 tok
- Min Mac RAM
- 8 GB
Qwen3 4B Instruct 2507 (4-bit)
Top all-rounder — reasoning, web search, PDF tools; passed every test
- RAM peak
- 3.3–4.1 GB
- Weights
- 2.12 GB
- Context
- 262,144 tok
- Min Mac RAM
- 8 GB
Qwen3 8B (4-bit)
Strongest reasoning in class — passed every on-device test
- RAM peak
- 5.5–6.3 GB
- Weights
- 4.31 GB
- Context
- 40,960 tok
- Min Mac RAM
- 16 GB
LFM (Liquid AI) 1
LFM2.5 1.2B Instruct (4-bit)
Fast, dependable chat and PDF/doc Q&A (Liquid AI)
- RAM peak
- 1.8–2.6 GB
- Weights
- 0.62 GB
- Context
- 128,000 tok
- Min Mac RAM
- 8 GB
SmolLM 1
SmolLM3 3B (4-bit)
Reasoning, PDFs, and doc lookup — passed every on-device test
- RAM peak
- 2.9–3.7 GB
- Weights
- 1.67 GB
- Context
- 65,536 tok
- Min Mac RAM
- 8 GB
Nemotron 3
Nemotron 3 Nano 4B (4-bit)
Reasoning, web search, and doc Q&A — near-perfect in tests
- RAM peak
- 3.2–4.0 GB
- Weights
- 2.24 GB
- Context
- 262,144 tok
- Min Mac RAM
- 8 GB
Llama Nemotron 8B UltraLong 1M (4-bit)
Long-context reasoning, web search, and doc Q&A on Mac
- RAM peak
- 5.5–6.3 GB
- Weights
- 4.52 GB
- Context
- 1,073,152 tok
- Min Mac RAM
- 16 GB
Nemotron Nano 9B v2 (4-bit)
Reasoning, web search, and docs — passed every on-device test
- RAM peak
- 6.0–6.8 GB
- Weights
- 5.0 GB
- Context
- 131,072 tok
- Min Mac RAM
- 16 GB
Ministral 1
Ministral 3 3B Instruct (4-bit)
Compact instruction-following chat and doc Q&A
- RAM peak
- 3.8–4.6 GB
- Weights
- 2.59 GB
- Context
- 262,144 tok
- Min Mac RAM
- 8 GB
Phi 1
Phi-4 (3-bit)
High-accuracy reasoning and math on Mac
- RAM peak
- 7.0–7.8 GB
- Weights
- 5.98 GB
- Context
- 16,384 tok
- Min Mac RAM
- 16 GB
Standings describe how each conversion behaves inside Mortise — not a general benchmark of the underlying model. How this was measured
// sovereign_vault_protocol
Enter the Sovereign Vault
Mortise is coming soon. No account required. Nothing leaves your device.