MODELS / PERFORMANCE
Choose models for your computer and workload. File size, estimated memory and measured speed are shown separately.
Configuration sharing, speaker tools and acceleration describe the current development build. See Downloads for released installer versions; older versions may not support app links.
| Computer | ASR + writer | Resource mode | Evidence |
|---|---|---|---|
| Apple Silicon · 16 GB everyday | zipformer-smallqwen3-1.7b | balanced · 2048 tokens | Starting point · unmeasured |
| Apple Silicon · 16 GB · experimental MLX | zipformer-smallqwen3-1.7b-mlx-4bit | balanced · 2048 tokens | Starting point · unmeasured |
| Apple Silicon · 24 GB after-meeting | zipformer-smallqwen3-4b | quality · 4096 tokens | Starting point · unmeasured |
| Apple Silicon · 8 GB economy | zipformer-smallqwen3-0.6b | eco · 2048 tokens | Starting point · unmeasured |
| M5 · 24 GB · measured Metal writer | zipformer-smallqwen3-1.7b | balanced · 2048 tokens | Measurement linked |
| General CPU · 16 GB everyday | zipformer-smallqwen3-1.7b | balanced · 2048 tokens | Starting point · unmeasured |
| General CPU · from 8 GB | zipformer-smallqwen3-0.6b | eco · 2048 tokens | Starting point · unmeasured |
Check available memory as well as installed memory. Prioritize recording and transcription during meetings; larger writers are better suited to after-meeting work.
Metal accelerates existing GGUF models. MLX uses separate MLX weights and runtime. Neither backend sends inference to a cloud service.
MLX is experimental and needs a separately installed compatible runtime on Apple Silicon; it does not directly load GGUF.
The CPU backend remains available. Economy, balanced and quality modes control threads, context and job budgets. A budget is a scheduling check, not an OS-enforced memory cap.
macOS 27.0.1 · arm64 · 2026-10-24
| Backend / model | Generationtokens/s | First tokenms | Peak process RSSMiB | Runs |
|---|---|---|---|---|
| CPUqwen3-1.7b | 13.5 | 2213 | 2426 | 3 |
| METALqwen3-1.7b | 78.3 | 102 | 1431 | 3 |
Fixed test input; speed and latency are medians across runs. Model warm, prompt cache cold. RSS is sampled process memory, not total GPU memory or system power.
{
"method": {
"kind": "llama-server-completion",
"prompt_sha256": "d1073ca91137cbd8b8c777e44e3634b6d02a3f492eaf4805ec47f5f68c9654fb",
"server_sha256": "4bc6b545e77fc15471493503fe2d63a3e7dfee0a46e6735b2244c8f4eb934bf5",
"llama_cpp_revision": "5aeac37a35b44fde4bdffaec9159f2fcb253b622",
"app_commit": "5cadd45b7a78e2c6abec380dcc2f6a89d31d3b66",
"app_dirty": true,
"generation_tokens_requested": 128,
"repetitions": 3,
"cache": "prompt cache disabled; model warm after one unmeasured 16-token request",
"rss_sampling_interval_ms": 100,
"synthetic_input": true
},
"parameters": [
{
"backend": "cpu",
"model_id": "qwen3-1.7b",
"context_size": 2048,
"threads": 2,
"batch_size": 128,
"ubatch_size": 128,
"generated_tokens": 128,
"quantization": "Q4_K_M"
},
{
"backend": "metal",
"model_id": "qwen3-1.7b",
"context_size": 2048,
"threads": 2,
"batch_size": 128,
"ubatch_size": 128,
"generated_tokens": 128,
"quantization": "Q4_K_M"
}
],
"limitations": [
"Single machine; not a prediction for other M-series chips.",
"Generation speed is not transcription accuracy or meeting-summary quality.",
"RSS is a sampled process-resident-memory peak, not total unified/GPU allocation.",
"Power draw and thermal throttling were not measured.",
"128 output tokens are forced for comparability; this is a short synthetic benchmark."
]
}
Downloads are checked against pinned sizes and SHA-256. Downloading is online; subsequent inference is offline. Memory figures are estimates and exclude headroom for other apps and the OS.