Use cases › Where it runs

Run it on your own machine: a phone, a 16 GB laptop or a 64 GB Mac

Run it on a phone, a 16 GB laptop or a 64 GB Mac, with no server in the loop.

267 text models that fit at these constraints. Data observed 2026-10-25 06:54 UTC.

Showing models for: 16 GB at INT4.

What it looks like in practice

The most-used model at each memory step, from Qwen3-0.6B at about 0.5 GB to Qwen3.8-Flash-Next at about 108 GB.

INT4 weights against a 16 GB card and a 64 GB Mac. Each row is the most downloaded model that needs more than the step below it.

Source: Estimate: weights only (parameters × 0.5 bytes × 1.2), KV cache extra; downloads from Hugging Face, observed 2026-10-25

Start with these four

One pick per constraint group: most downloaded at your memory among models that use at least a quarter of it, highest third-party tool-calling score at your memory, and the most used one size down and one size up.

Most used at 16 GB

Qwen3-14B

Qwen

Most downloaded in the last 30 days among text models with a GGUF build whose INT4 weights fit 16 GB and use at least a quarter of it (4 GB or more): 10M (Hugging Face, 2026-10-25).

~8.9 GBINT4 est.

10Mdownloads 30d

12GGUF builds

Sources: estimate method 2026-10-25 · huggingface.co 2026-10-25 · huggingface.co 2026-10-25

Best measured at 16 GB

Llama-xLAM-2-8b-fc-r

Salesforce

Highest third-party BFCL tool-calling score among models that fit 16 GB at INT4: 46.68% (gorilla.cs.berkeley.edu, 2026-10-25).

46.68%BFCL

~4.8 GBINT4 est.

91.9kdownloads 30d

Sources: gorilla.cs.berkeley.edu 2026-10-25 · estimate method 2026-10-25 · huggingface.co 2026-10-25

Fits 8 GB

Qwen3-0.6B

Qwen

Most downloaded in the last 30 days among models whose INT4 weights fit 8 GB: 24.9M (Hugging Face, 2026-10-25).

~0.5 GBINT4 est.

24.9Mdownloads 30d

7GGUF builds

Sources: estimate method 2026-10-25 · huggingface.co 2026-10-25 · huggingface.co 2026-10-25

64 GB Mac

gemma-4-31B-it

google

Most downloaded in the last 30 days among models too big for 16 GB that fit a 64 GB Mac at INT4: 8.5M (Hugging Face, 2026-10-25).

~19 GBINT4 est.

8.5Mdownloads 30d

10GGUF builds

Sources: estimate method 2026-10-25 · huggingface.co 2026-10-25 · huggingface.co 2026-10-25

Everything that fits 267

The first ten text models that fit, in the scoreboard's order for this job. The full list, sortable, lives in the scoreboard with these constraints.

Everything that fits: first 10 of 267
#				GGUF · MLX
1	Qwen3-0.6BQwen	752M	~0.5 GB	GGUF · MLX	40K	24.9M	Apache 2.0
2	Qwen3-4BQwen	4.02B	~2.4 GB	GGUF · MLX	40K	11.4M	Apache 2.0
3	Qwen3-14BQwen	14.8B	~8.9 GB	GGUF · MLX	40K	10M	Apache 2.0
4	gemma-4-26B-A4B-itgoogle	25.8B 4B active	~16 GB	GGUF · MLX	256K	8.6M	Apache 2.0
5	Qwen3-8BQwen	8.19B	~4.9 GB	GGUF · MLX	40K	8.2M	Apache 2.0
6	Qwen3.5-4BQwen	4.66B	~2.8 GB	GGUF · MLX	256K	7.5M	Apache 2.0
7	Qwen3.5-9BQwen	9.65B	~5.8 GB	GGUF · MLX	256K	6.3M	Apache 2.0
8	Qwen2.5-7B-InstructQwen	7.62B	~4.6 GB	GGUF	32K	6.1M	Apache 2.0
9	gpt-oss-20bopenai	20.9B	~13 GB	GGUF · MLX	128K	5.8M	Apache 2.0
10	Qwen2.5-0.5B-InstructQwen	494M	~0.3 GB	GGUF · MLX	32K	5.3M	Apache 2.0

Show all 267 in the scoreboard →

Most-shared on this job

Links about these models ranked by how many newsletters and link sources shared them in the last 28 days; each source counts once.All links

Shared by 1
Carbon-X (Hugging Face): A 14B DNA-plus-text model, midtrained from Qwen3-14B on 300B interleaved tokens, then trained…Latent Space / AINewsAbout Qwen3-14B
x.com · first seen 2026-10-23
Shared by 1
Early commenters dispute the published benchmark picture: one reports Flint is worse than Ornith-1.5-35B and “far worse…Latent Space / AINewsAbout Ornith-1.5-9B (series match)
github.com · first seen 2026-10-23
Shared by 1
Early commenters dispute the published benchmark picture: one reports Flint is worse than Ornith-1.5-35B and “far worse…Latent Space / AINewsAbout Ornith-1.5-9B (series match)
huggingface.co · first seen 2026-10-23
Shared by 1
Ant Group released Ming-3.0-tts-0.5B, a dense MIT-licensed TTS model built on Qwen2.5-0.5B.AlphaSignalAbout Qwen2.5-0.5B
huggingface.co · first seen 2026-10-22
Shared by 1
Official 4-bit AWQ build of Qwen2.5-3B-Instruct, 3.09B params, 32K context, 8K generationAlphaSignalAbout Qwen2.5-3B-Instruct
huggingface.co · first seen 2026-10-22

How people judge this job

How the memory estimate works

None of them measure speed on your own machine. The memory figures are weight estimates (parameters × bytes × 1.2 overhead); the KV cache for long contexts comes on top, and a Mac shares its unified memory with everything else you run.

Nearest scoreboard view: Runs on my machine. Follow changes to this job's evidence: RSS.