Which model should I run?

Tiny and quick — good for short replies when your phone is low on memory.

0.6B ~400–500 MB Phone 3 GB+ Mac 8 GB+ GPU Any GGUF · MLX

Qwen3-0.6B; Q4 GGUF & 4-bit MLX. Hugging Face downloads in the tens of millions.

2026 Qwen tiny — hybrid long-context brain in a phone-sized file. Best when you want the newest Qwen on 4–6 GB devices.

0.8B ~500–600 MB Phone 3 GB+ Mac 8 GB+ GPU Any GGUF · MLX

Qwen3.5-0.8B (Feb 2026). Gated DeltaNet hybrid; 262K native context. Community GGUF/MLX as engines catch up — import in LM Mini or host via Connect.

Google’s smallest Gemma 3 chat model. Great starter for everyday questions.

1B ~700–750 MB Phone 3 GB+ Mac 8 GB+ GPU Any GGUF · MLX

google/gemma-3-1b-it. Instruct-tuned; GGUF + MLX.

Google’s 2026 edge Gemma — ~2.3B effective, text + image + audio. The new tiny phone default if your engine supports Gemma 4.

2.3B ~1.4–1.8 GB (Q4/QAT) Phone 4 GB+ Mac 8 GB+ GPU Any GGUF

google/gemma-4-E2B-it (Apr 2026). Official QAT GGUF. Apache 2.0. ~5.1B with embeddings; plan ~2–3 GB live.

Meta’s pocket Llama — fast, tool-friendly, and solid for light chat.

1B ~700–808 MB Phone 3 GB+ Mac 8 GB+ GPU Any GGUF · MLX

Llama-3.2-1B-Instruct; tool calling. GGUF + MLX.

Hugging Face’s small all-rounder. Surprisingly capable for its size.

1.7B ~1.0 GB Phone 3 GB+ Mac 8 GB+ GPU Any GGUF

SmolLM2-1.7B-Instruct Q4_K_M via LM Studio / custom HF import.

A mini “thinking” model — stronger at math and step-by-step reasoning.

1.5B ~1.0–1.1 GB Phone 3 GB+ Mac 8 GB+ GPU Any GGUF · MLX

R1 distill of Qwen 2.5 1.5B; reasoning traces. GGUF + MLX.

Best everyday phone pick for most people — chat, tools, and general help.

1.7B ~1.1–1.2 GB Phone 3 GB+ Mac 8 GB+ GPU Any GGUF · MLX

Qwen3-1.7B; primary free-slot family in LM Mini. GGUF + MLX.

Newest small Qwen that still fits everyday phones — sharper than 1.7B-class, still downloadable on cellular.

2B ~1.3–1.5 GB Phone 4 GB+ Mac 8 GB+ GPU Any GGUF · MLX

Qwen3.5-2B. Hybrid attention + native multimodal in the 3.5 family. Prefer Q4 GGUF once your llama.cpp/MLX build lists Qwen3.5.

Previous-gen Qwen that’s still a reliable tiny assistant.

1.5B ~1.0 GB Phone 3 GB+ Mac 8 GB+ GPU Any GGUF

Qwen2.5-1.5B-Instruct; still huge Hugging Face download volume.

Google’s on-device Gemma 3n (effective 2B). Built for phones, including audio-aware builds.

2B ~1.5 GB Phone 4 GB+ Mac 8 GB+ GPU Any GGUF

google/gemma-3n-E2B-it. LiteRT / GGUF community quants. Newer than Gemma 3 1B.

Noticeably smarter than 1B — good for writing help and longer chats on Pro phones.

3B ~1.9–2.0 GB Phone 6 GB+ Mac 8–16 GB GPU 6 GB+ VRAM GGUF · MLX

Llama-3.2-3B-Instruct; tool calling. GGUF + MLX.

Hugging Face’s 2025 small model — a step up from SmolLM2, still phone-sized.

3B ~1.9 GB (Q4) Phone 6 GB+ Mac 8–16 GB GPU 6 GB+ VRAM GGUF

HuggingFaceTB/SmolLM3-3B. GGUF via bartowski / community.

Google’s mid-size Gemma 3 — sharper answers; some builds can look at images.

4B ~2.7–3.7 GB Phone 6 GB+ Mac 8–16 GB GPU 8 GB+ VRAM GGUF · MLX

google/gemma-3-4b-it; multimodal on supported builds. GGUF + MLX.

2026’s 4B daily driver — long context and stronger tools than Qwen 3 4B, if your phone has ~6 GB free.

4B ~2.5–2.8 GB Phone 6 GB+ Mac 8–16 GB GPU 8 GB+ VRAM GGUF · MLX

Qwen3.5-4B; millions of HF downloads. Hybrid Gated DeltaNet. Import a Q4 when the engine supports it; Qwen 3 4B Instruct 2507 remains the in-app pick today.

Google’s on-device Gemma 4 (~4.5B effective) with image and audio. The 2026 upgrade from Gemma 3 4B / 3n E4B.

4.5B ~2.6–3.2 GB (Q4/QAT) Phone 8 GB+ Mac 8–16 GB GPU 8 GB+ VRAM GGUF

google/gemma-4-E4B-it. Official QAT GGUF. ~8B with embeddings; budget ~4 GB live. Apache 2.0.

Strong reasoning in a phone-friendly package — great for explanations and code-ish help.

3.8B ~2.3–2.4 GB Phone 6 GB+ Mac 8–16 GB GPU 8 GB+ VRAM GGUF · MLX

microsoft/Phi-4-mini-instruct (~3.8B). GGUF + MLX.

Balanced daily driver for Pro phones and Macs — quality without a desktop GPU.

4B ~2.4–2.5 GB Phone 6 GB+ Mac 8–16 GB GPU 8 GB+ VRAM GGUF · MLX

July 2025 Instruct refresh; millions of HF downloads. GGUF + MLX.

Compact Qwen for chat and light coding when 4B feels heavy.

3B ~1.9 GB Phone 6 GB+ Mac 8–16 GB GPU 6 GB+ VRAM GGUF

Qwen2.5-3B-Instruct Q4_K_M; LM Studio / Ollama.

Gemma 3n effective-4B — Google’s newer on-device stack, including multimodal 3n builds.

4B ~2.6 GB (Q4) Phone 8 GB+ Mac 8–16 GB GPU 8 GB+ VRAM GGUF

google/gemma-3n-E4B-it. Community GGUF (bartowski). Heavier than E2B.

NVIDIA’s 2026 tiny Nemotron — dense 4B-class, aimed at local agents.

4B ~2.5 GB (Q4) Phone 8 GB+ Mac 16 GB+ GPU 8 GB+ VRAM GGUF

nvidia/NVIDIA-Nemotron-3-Nano-4B-BF16. Use a GGUF quant on phone; BF16 is a desktop download.

Classic open model — clear writing and solid general knowledge on Mac or GPU.

7B ~4.1 GB (Q4) Phone 12 GB+ Mac 16 GB+ GPU 8 GB+ VRAM GGUF

Mistral-7B-Instruct v0.3; ubiquitous GGUF quants.

Meta’s workhorse 8B — excellent general assistant on Mac or a home GPU.

8B ~4.7 GB (Q4) Phone 12 GB+ Mac 16 GB+ GPU 8–12 GB VRAM GGUF · MLX

Llama-3.1-8B-Instruct; tool calling; LM Studio / Ollama staple.

Frontier feel on-device for high-RAM phones and M-series Macs.

8B ~4.7–4.9 GB Phone 12 GB+ Mac 16 GB+ GPU 10–12 GB VRAM GGUF · MLX

Qwen3-8B; tool calling. GGUF + MLX in LM Mini.

Frontier feel on a 16 GB phone or any M-series Mac — the 2026 8B-class upgrade.

9B ~5.2–5.6 GB (Q4) Phone 16 GB+ Mac 18–24 GB GPU 12 GB+ VRAM GGUF · MLX

Qwen3.5-9B (most-downloaded 3.5 size on HF). Not in the LM Mini in-app list yet; LM Studio / Ollama / custom import + Connect.

Popular all-rounder for coding help, summaries, and chat on desktop or Mac.

7B ~4.4 GB (Q4) Phone 12 GB+ Mac 16 GB+ GPU 8–10 GB VRAM GGUF · MLX

Qwen2.5-7B-Instruct — still one of the most downloaded instruct models on HF.

Reasoning-focused — walks through hard problems before answering.

8B ~4.7 GB (Q4) Phone 12 GB+ Mac 16 GB+ GPU 10–12 GB VRAM GGUF

R1 distill on Llama 8B; longer CoT traces.

Allen AI’s fully open 7B instruct model — a transparent alternative to Llama/Qwen.

7B ~4.3 GB (Q4) Phone 12 GB+ Mac 16 GB+ GPU 8–10 GB VRAM GGUF

allenai/Olmo-3-7B-Instruct (2025). Community GGUF.

NVIDIA 9B local model — stronger than 8B class if you have the RAM.

9B ~5.2 GB (Q4) Phone 16 GB+ Mac 18–24 GB GPU 12 GB+ VRAM GGUF

nvidia/NVIDIA-Nemotron-Nano-9B-v2. Prefer Q4 GGUF on Mac/GPU.

Google’s 9B chat model — polished answers when you have Mac RAM or a GPU.

9B ~5.4 GB (Q4) Phone 16 GB+ Mac 18–24 GB GPU 12 GB+ VRAM GGUF · MLX

Gemma-2-9B-IT; still excellent instruction following.

Community favorite for creative writing and agent-style tool use.

8B ~4.7 GB (Q4) Phone 12 GB+ Mac 16 GB+ GPU 10–12 GB VRAM GGUF

NousResearch Hermes-3-Llama-3.1-8B; function calling.

Mixture-of-experts: 30B total, ~3B active. Big-model taste on a strong Mac or GPU.

30B ~16 GB (Q4 MoE) Phone — Mac 32 GB+ GPU 16–24 GB VRAM GGUF

Qwen3-30B-A3B-Instruct-2507 MoE. Q4 is lighter than dense 30B; still not a phone model.

Dense 2026 Qwen for a high-RAM Mac or a 24 GB GPU — then use the phone as the remote keyboard.

27B ~16 GB (Q4) Phone — Mac 32 GB+ GPU 24 GB+ VRAM GGUF

Qwen3.5-27B. Q4 ~16 GB. Also Qwen3.6-27B as a later dense refresh — same RAM class.

Large Gemma 3 with vision — Mac-first in LM Mini (4-bit MLX), or a home GPU via Connect.

27B ~15.5 GB (4-bit) Phone — Mac 24–36 GB GPU 24 GB+ VRAM MLX · GGUF

mlx-community/gemma-3-27b-it-4bit in LM Mini (~15.5 GB). Needs ~24 GB unified memory.

Mac-first Gemma 3 — premium quality when you have unified memory to spare.

12B ~7.1 GB Phone — Mac 24 GB+ GPU 16 GB+ VRAM MLX · GGUF

google/gemma-3-12b-it. MLX 4-bit in LM Mini on Mac.

2026 Gemma 4 12B — text, image, and audio. Mac or GPU; too big for phones.

12B ~7–8 GB (Q4/QAT) Phone — Mac 24 GB+ GPU 16 GB+ VRAM GGUF

google/gemma-4-12B-it. Official QAT GGUF. 256K context. Host in LM Studio then chat from LM Mini.

Desktop-quality Qwen 3 in LM Mini on Mac — 4-bit MLX on 16 GB+ unified memory.

14B ~8.2 GB (4-bit) Phone — Mac 16–32 GB GPU 12–16 GB VRAM MLX · GGUF

mlx-community/Qwen3-14B-4bit in LM Mini (~8.2 GB). GGUF Q4 similar. Phone: Connect only.

Serious desktop quality — best on a home GPU or a high-RAM Mac.

14B ~8–9 GB (Q4) Phone — Mac 32 GB+ GPU 12–16 GB VRAM GGUF · MLX

Qwen2.5-14B-Instruct Q4/Q5; LM Studio / Ollama.

Heavy reasoning model for research-style answers on GPU or big Macs.

14B ~8–9 GB (Q4) Phone — Mac 32 GB+ GPU 16 GB+ VRAM GGUF

R1 14B distill; long context CoT; prefer desktop.

2025 Mistral Small — rich writing and analysis. GPU or 48 GB+ Mac.

24B ~13 GB (Q4) Phone — Mac 48 GB+ GPU 16–24 GB VRAM GGUF

Mistral-Small-3.2-24B-Instruct-2506. Q4 ~13 GB.

Near-frontier open weights for a well-equipped GPU PC.

32B ~18 GB (Q4) Phone — Mac 64 GB+ GPU 24 GB+ VRAM GGUF

Qwen2.5-32B-Instruct; Q4 ~18 GB; LM Studio remote.

Top in-app Qwen 3 on 32 GB+ Macs (4-bit MLX). Phones should Connect, not download.

32B ~18 GB (4-bit) Phone — Mac 32–64 GB GPU 24 GB+ VRAM MLX · GGUF

mlx-community/Qwen3-32B-4bit in LM Mini (~18 GB). minRam 32 GB.

Meta’s Llama 4 Scout MoE (17B×16E). Home GPU / big Mac — then chat from the phone via Connect.

109B ~40 GB+ (Q4 MoE) Phone — Mac — (use remote) GPU 48 GB+ VRAM GGUF

Llama-4-Scout-17B-16E-Instruct. Mixture-of-experts; Q4 still tens of GB.

Meta’s larger Llama 4 MoE (17B×128E). Multi-GPU / 80 GB class — Connect from the phone; do not download to the handset.

400B ~80 GB+ (Q4 MoE) Phone — Mac — (use remote) GPU 80 GB+ VRAM GGUF

Llama-4-Maverick-17B-128E-Instruct. Far heavier than Scout. FP8 and GGUF community quants.

MoE Gemma 4: 26B total, ~4B active. Big-model taste at 8B-class compute — GPU or 32 GB+ Mac, then Connect.

26B ~8–12 GB (Q4 MoE) Phone — Mac 32 GB+ GPU 16 GB+ VRAM GGUF

google/gemma-4-26B-A4B-it. Official QAT GGUF. Efficiency king of the Gemma 4 stack.

Dense Gemma 4 flagship — reasoning, vision, coding. Home GPU / 48 GB Mac, chat from the phone via Connect.

31B ~18 GB (Q4/QAT) Phone — Mac 48 GB+ GPU 24 GB+ VRAM GGUF

google/gemma-4-31B-it. Official QAT GGUF. 256K context. Apache 2.0.

2026 Qwen MoE: 35B total, ~3B active. Strong Mac/GPU pick — not a phone download.

35B ~12–16 GB (Q4 MoE) Phone — Mac 32 GB+ GPU 16–24 GB VRAM GGUF

Qwen3.5-35B-A3B. Lighter than dense 27B at similar quality. Qwen3.6-35B-A3B is the later FP8/MoE refresh in the same RAM band.

Top-tier Meta 70B — run at home on a strong GPU, chat from your phone via Connect.

70B ~40 GB (Q4) Phone — Mac — (use remote) GPU 48 GB+ VRAM GGUF

Llama-3.3-70B-Instruct; Q4 ~40 GB; multi-GPU friendly.

Huge Qwen for maximum quality on a serious home or office GPU rig.

72B ~40+ GB (Q4) Phone — Mac — (use remote) GPU 48 GB+ VRAM GGUF

Qwen2.5-72B-Instruct; Q4 ~40+ GB; Ollama / LM Studio.

Research-grade reasoning — pair with Connect so your phone stays light.

70B ~40+ GB (Q4) Phone — Mac — (use remote) GPU 48 GB+ VRAM GGUF

70B R1 distill; expects large VRAM or multi-GPU.

Mixture-of-experts — high quality without always loading a dense 70B.

46.7B ~26 GB (Q4) / less active Phone — Mac 64 GB+ GPU 24 GB+ VRAM GGUF

Sparse MoE; ~12–26 GB depending on quant; LM Studio classic.