Local AI benchmarks for Mac (2026): Outlier vs Ollama vs LM Studio
- A small chat model, Outlier Nano 4B, runs at ~70 tok/s on an M1 Ultra. We have not run Ollama or LM Studio on the same Mac, so this page makes no like-for-like comparison.
- For 27B coding, Outlier Core 27B does 20.7 tok/s and is the strongest coding tier in the lineup.
- 397B-class MoE? Outlier's paged engine runs it on a 64 GB Mac (V9 paged engine; ~1.59 tok/s measured on an M1 Ultra with 64 GB). We have not tested other apps on this model.
- Accuracy is on the tier benchmarks page: MMLU for every tier (Nano 52.0%, n=200, build 1.11.804) and HumanEval for Core 27B and Vision 3.8. This page sticks to tok/s and memory.
Outlier Pro: $49 once · Black Friday price, all of November: $24.50 once, until 30 November
Free to start (Nano and Lite). macOS 15+, Apple silicon. Refund window: 30 days, no questions asked.
Raw numbers, not vibes. This is the benchmark data behind every major local-AI option on Apple Silicon in 2026. Outlier's figures come off my dev M1 Ultra. Ollama and LM Studio numbers cite public benchmark posts where it makes sense.
Outlier lineup: measured (M1 Ultra, MLX 4-bit, batch 1, 4096 prefill, 256 decode)
| Tier | Params | Disk | RAM (peak) | Decode tok/s |
|---|---|---|---|---|
| Nano 4B | 4B dense | ~2.4 GB | ~4 GB | 71.7 |
| Lite 9B | 9B dense | ~5 GB | ~7 GB | 53.4 |
| Quick 26B | 26B-a4b MoE | ~16 GB | ~17 GB | 14.6 . Re-measured 43.6 on 2026-08-24 (v1.11.804); the 14.6 is what this 2026-04-29 bench recorded and is left as run |
| Core 27B | 27B dense | ~15 GB | ~17 GB | 20.7 |
| Vision 35B-A3B: this 2026-04-29 run is Vision 3.6; the shipping tier is now Vision 3.8 (Qwen3.8-27B, 15.5 GB) | 35B-A3B MoE | ~20 GB | ~18 GB (cap) / ~3.5 GB (V10) | ~8 (V10) / 16 (cap) |
| Plus 397B-A17B | 397B-A17B MoE | ~209 GB | ~14 GB (V9 paged) | 1.59 (V9): the pre-ship K-override sweep at the shipped K=20, on an M1 Ultra with 64 GB (5 prompts × 80 tokens, fresh process, model load excluded; per-prompt 1.35/1.52/1.62/1.65/1.64). The engine-comparison page reports 2.1 for the same tier under its own per-token RSS protocol. Where they disagree we publish the lower, shipping-config figure, and this page's dataset.json does the same. |
Source: Outlier FINAL_LAUNCH_NUMBERS.md. The M1 Ultra Mac tok/s bench ran 2026-04-29. Accuracy figures are on the tier benchmarks page (MMLU n=200, build 1.11.804, 2026-08-24).
Outlier decode speed by tier (measured)
| Tool | Model | Format | Mac | Decode tok/s |
|---|---|---|---|---|
| Outlier | Nano 4B | MLX 4-bit | M1 Ultra 64 GB | 71.7 |
| Outlier | Nano 4B | MLX 4-bit | M4 Air 16 GB | ~32 |
| Outlier | Core 27B | MLX 4-bit | M1 Ultra 64 GB | 20.7 |
| Outlier | Plus 397B-A17B | MLX 4-bit, V9 paged | M1 Ultra 64 GB | 1.59 |
Source: Outlier numbers measured locally on an M1 Ultra (the M4 Air figure is a single-prompt observation, 2026-05-04, indicative). We have not run Ollama or LM Studio on the same Macs, so this page makes no like-for-like comparison.
Cost comparison (24 months, single Mac developer)
| Setup | 24-month total |
|---|---|
| Outlier Free (Nano + Lite) | $0 |
| Outlier Lifetime Licence ($49 once, lifetime Pro) | $49 |
| Ollama (OSS, free) | $0 |
| LM Studio (free for personal use) | $0 |
| ChatGPT Plus / Claude Pro ($20/mo) | $480 billed monthly; Claude’s annual rate ($200 a year) makes it $400 (claude.com/pricing, as of October 2026; ChatGPT Plus per openai.com/chatgpt/pricing, read 2026-09-08) |
| Claude Max (from $100/mo) | from $2,400 (claude.com/pricing, as of October 2026) |
Corrected 2026-09-08. This table carried two Outlier rows that no longer describe anything we sell: a $216 24‑month total, which only adds up if a $9/month plan exists, Outlier has no monthly plan, and a $298 one-time price that contradicted the price row beneath it. Both removed; $49 once is the only Outlier price. Rival figures re-read the same day: Claude Pro is $20 monthly or $17/month annually, and ChatGPT Pro and Claude Max both start at $100/month, not $200.
What the bench doesn't measure
- Prompt prefill / TTFT. Decode tok/s is the steady-state figure. Feed a long prompt and your first token can still be several seconds out, even on the fast tiers.
- Tier swap cost. Switch to a tier you haven't touched in a while and the cold load runs 10s–74s, scaling with size.
- Quality on agentic loops. The bench tested single-turn outputs. Multi-turn agent quality swings more on the app than the model.
- Vision tasks rigorously. Those Vision 35B numbers are language-only. A real image-task quality bench is still coming.
- Long-context (50k+). Every bench prompt was capped at 4096 prefill. Push past that and the cloud flagships still pull ahead by a real margin.
How to reproduce
- Outlier: grab it from outlier.host, then run the bundled benchmark harness in the app's developer console.
- Ollama:
ollama run <model> --verboseprints tok/s for every response. - LM Studio: the chat REPL shows tok/s right in the response footer.
- Accuracy evals: reach for lm-evaluation-harness, loading each model through its native backend.
Questions about these Mac benchmarks
How fast is Outlier on a Mac?
On M1 Ultra: Nano 4B 71.7 tok/s, Lite 9B 53.4, Core 27B 20.7, Plus 397B 1.59 on the V9 paged engine. Nano hit about 32 tok/s on an M4 Air in a single prompt (2026-05-04, indicative).
How does Outlier compare to Ollama and LM Studio?
This page publishes no head-to-head tok/s for them, so it makes no speed claim either way; Outlier uniquely runs MoE models bigger than RAM via its V9 paged engine.
What are Outlier's accuracy numbers?
Core 27B is the reasoning-strong coding tier; Nano 4B is the fast everyday tier. For accuracy, see the tier benchmarks page: Nano scored 52.0% on a 200-question MMLU sample and Core 27B 89.5% (build 1.11.804, M1 Ultra).
Try Outlier free
Free: Nano + Lite. Pro: $49 once; Black Friday price, all of November: $24.50 once, until 30 November. macOS 15+. Refund window: 30 days, no questions asked.
Download free (276 MB) Buy ProOn your phone?