Twoody
Local LLM calculator
Which local LLM can your computer run — and how fast?
Pick your Mac or PC. See which open-weight models fit its memory, how fast they write, how long they take to read a document, and what they are good for.
The models for this computer
On a PC with GeForce RTX 3070 (8 GB) and 32 GB of RAM: 23 of 62 models run well.
What Twoody suggests for this PC
On Windows as on Linux: Qwen3.5 9B, in one click, with conversations of 64k tokens.
From 24 GB, Twoody also offers the larger Qwen3.6 35B-A3B: it needs 16.8 GB of memory on this computer.
Twoody for Windows and Linux is in private beta: on Windows, it runs the model on the graphics card through Vulkan when the driver supports it, on the processor otherwise; on Linux, on the processor. The speeds below are estimates for a model running on the graphics card, not measurements made with Twoody.
Run well on this computer
6.7 of 6.6 GB
Runs — close other apps
41–59 tokens/s
feels instant
20 pages in 11 s
- Chat and writing
- Your documents
- Translation
7.2 of 6.6 GB
Runs — close other apps
48–69 tokens/s
feels instant
20 pages in 9 s
- Chat and writing
- Your documents
- Translation
- Reasoning
6.9 of 6.6 GB
Runs — close other apps
50–72 tokens/s
feels instant
20 pages in 9 s
- Chat and writing
- Your documents
- Translation
6.6 of 6.6 GB
Runs — close other apps
71–103 tokens/s
feels instant
20 pages in 5 s
- Chat and writing
- Translation
4.9 of 6.6 GB
Runs comfortably
62–123 tokens/s
feels instant
20 pages in 5 s
- Chat and writing
- Translation
4.9 of 6.6 GB
Runs comfortably
80–116 tokens/s
feels instant
20 pages in 4 s
- Chat and writing
- Translation
3.6 of 6.6 GB
Runs comfortably
56–113 tokens/s
feels instant
20 pages in 5 s
- Chat and writing
5 of 6.6 GB
Runs comfortably
99–142 tokens/s
feels instant
20 pages in 4 s
- Chat and writing
- Translation
4.6 of 6.6 GB
Runs comfortably
84–120 tokens/s
feels instant
20 pages in 4 s
- Chat and writing
- Translation
3.9 of 6.6 GB
Runs comfortably
100+ tokens/s
feels instant
20 pages in 4 s
- Chat and writing
- Translation
4.7 of 6.6 GB
Runs comfortably
100+ tokens/s
feels instant
20 pages in 3 s
2.9 of 6.6 GB
Runs comfortably
100+ tokens/s
feels instant
20 pages in 3 s
1.7 of 6.6 GB
Runs comfortably
100+ tokens/s
feels instant
20 pages in 1 s
6.7 of 6.6 GB
Runs — close other apps
41–59 tokens/s
feels instant
20 pages in 11 s
- Chat and writing
- Your documents
- Translation
- Reasoning
7.2 of 6.6 GB
Runs — close other apps
48–69 tokens/s
feels instant
20 pages in 9 s
- Chat and writing
- Your documents
- Translation
- Reasoning
6.9 of 6.6 GB
Runs — close other apps
50–72 tokens/s
feels instant
20 pages in 9 s
- Chat and writing
- Your documents
- Translation
- Reasoning
5.5 of 6.6 GB
Runs comfortably
54–78 tokens/s
feels instant
20 pages in 9 s
- Chat and writing
- Your documents
- Translation
- Reasoning
4.9 of 6.6 GB
Runs comfortably
80–116 tokens/s
feels instant
20 pages in 4 s
- Chat and writing
- Translation
4.6 of 6.6 GB
Runs comfortably
84–120 tokens/s
feels instant
20 pages in 4 s
- Chat and writing
3.2 of 6.6 GB
Runs comfortably
97–140 tokens/s
feels instant
20 pages in 3 s
- Chat and writing
- Translation
2.4 of 6.6 GB
Runs comfortably
89–178 tokens/s
feels instant
20 pages in 3 s
2.5 of 6.6 GB
Runs comfortably
100+ tokens/s
feels instant
20 pages in 3 s
1.6 of 6.6 GB
Runs comfortably
100+ tokens/s
feels instant
20 pages in 1 s
Run, but much slower (33)
16.8 of 6.6 GB
Partly on the processor: much slower
16–33 tokens/s
near instant
20 pages in 17 s
- Chat and writing
- Your documents
- Translation
- Code
- Reasoning
- Agents
21.2 of 6.6 GB
Partly on the processor: much slower
12–24 tokens/s
faster than you read
20 pages in 17 s
- Chat and writing
- Your documents
- Translation
- Code
- Reasoning
- Agents
22.9 of 6.6 GB
Partly on the processor: much slower
11–21 tokens/s
faster than you read
20 pages in 17 s
- Chat and writing
- Your documents
- Translation
- Code
- Reasoning
- Agents
22.7 of 6.6 GB
Partly on the processor: much slower
1.7–2.5 tokens/s
slower than you read
20 pages in 2 min
- Chat and writing
- Your documents
- Translation
- Code
- Reasoning
- Agents
19.6 of 6.6 GB
Partly on the processor: much slower
1.9–2.7 tokens/s
slower than you read
20 pages in 2 min
- Chat and writing
- Your documents
- Translation
- Code
- Agents
21.4 of 6.6 GB
Partly on the processor: much slower
8.2–16 tokens/s
faster than you read
20 pages in 20 s
- Chat and writing
- Your documents
- Translation
- Code
- Reasoning
- Agents
19.4 of 6.6 GB
Partly on the processor: much slower
2–2.9 tokens/s
slower than you read
20 pages in 88 s
- Chat and writing
- Your documents
- Translation
- Code
- Reasoning
- Agents
19.2 of 6.6 GB
Partly on the processor: much slower
12–21 tokens/s
faster than you read
20 pages in 22 s
- Chat and writing
- Your documents
- Translation
- Code
- Reasoning
- Agents
20.5 of 6.6 GB
Partly on the processor: much slower
1.5–3.1 tokens/s
slower than you read
20 pages in 79 s
- Chat and writing
- Your documents
- Translation
- Code
- Reasoning
- Agents
21.4 of 6.6 GB
Partly on the processor: much slower
1.4–2.9 tokens/s
slower than you read
20 pages in 79 s
- Chat and writing
- Your documents
- Translation
- Code
- Reasoning
- Agents
21.5 of 6.6 GB
Partly on the processor: much slower
1.4–2.9 tokens/s
slower than you read
20 pages in 79 s
- Chat and writing
- Your documents
- Translation
- Code
- Reasoning
- Agents
15.5 of 6.6 GB
Partly on the processor: much slower
13–23 tokens/s
faster than you read
20 pages in 26 s
- Chat and writing
- Your documents
- Translation
- Code
- Reasoning
- Agents
17.3 of 6.6 GB
Partly on the processor: much slower
2.7–3.9 tokens/s
slower than you read
20 pages in 67 s
- Chat and writing
- Your documents
- Translation
- Code
- Agents
11.2 of 6.6 GB
Partly on the processor: much slower
6–8.6 tokens/s
about your reading speed
20 pages in 41 s
- Chat and writing
- Your documents
- Translation
- Reasoning
11.9 of 6.6 GB
Partly on the processor: much slower
5.5–8 tokens/s
about your reading speed
20 pages in 41 s
- Chat and writing
- Your documents
11.6 of 6.6 GB
Partly on the processor: much slower
6.2–9 tokens/s
about your reading speed
20 pages in 39 s
- Chat and writing
- Your documents
- Translation
7.6 of 6.6 GB
Partly on the processor: much slower
16–23 tokens/s
near instant
20 pages in 34 s
- Chat and writing
- Your documents
- Translation
- Reasoning
7.9 of 6.6 GB
Partly on the processor: much slower
17–24 tokens/s
near instant
20 pages in 29 s
- Chat and writing
- Your documents
- Translation
- Reasoning
7.9 of 6.6 GB
Partly on the processor: much slower
13–27 tokens/s
near instant
20 pages in 28 s
- Chat and writing
- Your documents
- Translation
- Reasoning
8.4 of 6.6 GB
Partly on the processor: much slower
16–24 tokens/s
near instant
20 pages in 25 s
- Chat and writing
- Your documents
- Translation
8 of 6.6 GB
Partly on the processor: much slower
17–25 tokens/s
near instant
20 pages in 25 s
- Chat and writing
- Your documents
- Translation
8.2 of 6.6 GB
Partly on the processor: much slower
19–27 tokens/s
near instant
20 pages in 20 s
- Chat and writing
- Your documents
22.8 of 6.6 GB
Partly on the processor: much slower
1.7–2.5 tokens/s
slower than you read
20 pages in 2 min
- Chat and writing
- Your documents
- Translation
- Code
- Reasoning
- Agents
18.1 of 6.6 GB
Partly on the processor: much slower
15–25 tokens/s
faster than you read
20 pages in 20 s
- Chat and writing
- Your documents
- Translation
- Code
- Reasoning
- Agents
19.1 of 6.6 GB
Partly on the processor: much slower
12–21 tokens/s
faster than you read
20 pages in 22 s
- Chat and writing
- Your documents
- Translation
- Code
- Agents
19.2 of 6.6 GB
Partly on the processor: much slower
12–21 tokens/s
faster than you read
20 pages in 22 s
- Chat and writing
- Your documents
- Translation
- Code
- Reasoning
- Agents
16.1 of 6.6 GB
Partly on the processor: much slower
2.8–4 tokens/s
slower than you read
20 pages in 67 s
- Chat and writing
- Your documents
- Translation
- Code
- Reasoning
- Agents
16.1 of 6.6 GB
Partly on the processor: much slower
2.8–4 tokens/s
slower than you read
20 pages in 67 s
- Chat and writing
- Your documents
- Translation
- Code
- Agents
12 of 6.6 GB
Partly on the processor: much slower
14–25 tokens/s
faster than you read
20 pages in 24 s
- Chat and writing
- Your documents
- Translation
- Code
- Reasoning
- Agents
10.4 of 6.6 GB
Partly on the processor: much slower
26–45 tokens/s
near instant
20 pages in 16 s
- Chat and writing
- Your documents
- Translation
- Code
11.7 of 6.6 GB
Partly on the processor: much slower
5.7–8.1 tokens/s
about your reading speed
20 pages in 41 s
- Chat and writing
- Your documents
- Translation
- Reasoning
11.9 of 6.6 GB
Partly on the processor: much slower
5.5–8 tokens/s
about your reading speed
20 pages in 41 s
- Chat and writing
- Your documents
- Reasoning
10.5 of 6.6 GB
Partly on the processor: much slower
7–10 tokens/s
about your reading speed
20 pages in 39 s
- Chat and writing
- Your documents
- Translation
- Reasoning
Too big for this computer (6)
- Llama 3.3 70B · 44.9 of 6.6 GB
- Qwen3-Next 80B-A3B · 46.3 of 6.6 GB · Not in Twoody's catalogue
- gpt-oss 120B · 59.9 of 6.6 GB · Not in Twoody's catalogue
- Mistral Small 4 119B · 69.4 of 6.6 GB · Not in Twoody's catalogue
- Qwen3.5 122B-A10B · 74.7 of 6.6 GB · Not in Twoody's catalogue
- Qwen3 235B-A22B Instruct 2507 · 136.1 of 6.6 GB · Not in Twoody's catalogue
Each model's memory is what Twoody shows in “Models” for this computer: the same formula, the same files (the one for reading images included), the same conversation length — catalogue checked on October 15, 2026. Speeds are estimates, not measurements made with Twoody, computed from public llama.cpp benchmarks. Figures reviewed on October 16, 2026.
How to read the results
Choose your computer and its memory. The calculator goes through the models of Twoody's catalogue, and a few larger ones it does not offer, and tells how each one would run on that computer.
- What Twoody suggests
- The model the app installs first on that computer, one it has measured with its memory, its tools and your documents, and the memory it needs there.
- Run well on this computer
- The models that fit its memory, the strongest first: “Runs comfortably”, or “Runs — close other apps” when little room is left. Further down, the models that would run partly on the processor, much slower, then those that are too big.
- Memory
- What Twoody shows in “Models” for that computer: the model's files, the conversation it keeps in mind and room to work. The bar shows how much it takes of the memory a model may use there.
- Writes
- How fast the model writes its answer, in tokens per second. A token is a piece of a word, and you read silently at about 5 tokens per second.
- Reads a document
- How long the model takes to read a document of about 20 pages before it answers about it.
- Good for
- What its size suits: chat and writing, your documents, translation, code, reasoning or agents.
- One click in Twoody
- The model is in Twoody's catalogue and installs from the app in one click; the others say “Not in Twoody's catalogue”.
How much memory do you need for a local LLM?
Memory decides which models fit: the model's file and the one that lets it read images, the conversation it keeps in mind (8k to 64k tokens depending on the computer's memory, as in Twoody), and a little room to work. A Mac lets its graphics use about two thirds of its memory up to 32 GB, three quarters above. What runs comfortably:
| Memory | Strongest that run comfortably | Suggested by Twoody by default |
|---|---|---|
| 8 GB | Qwen3.5 4B, Qwen3 4B Instruct 2507, Nemotron 3 Nano 4B | Qwen3.5 4B (4 GB) |
| 16 GB | Gemma 4 12B, GLM-4.6V Flash, Qwen3.5 9B | Qwen3.5 9B (7.4 GB) |
| 24 GB | Qwen3 14B, Phi-4 14B, Ministral 3 14B | Qwen3.5 9B (7.4 GB) |
| 32 GB | Qwen3.6 35B-A3B, Gemma 4 31B, Qwen3 30B-A3B Instruct 2507 | Qwen3.5 9B (7.9 GB) |
| 48 GB | Qwen3.6 35B-A3B, Qwen3.6 35B-A3B (4 bits), Qwen3.5 35B-A3B | Qwen3.5 9B (7.9 GB) |
| 64 GB | Qwen3-Next 80B-A3B*, Llama 3.3 70B, Qwen3.6 35B-A3B | Qwen3.5 9B (7.9 GB) |
| 128 GB | Qwen3.5 122B-A10B*, Mistral Small 4 119B*, Qwen3-Next 80B-A3B* | Qwen3.5 9B (7.9 GB) |
| 256 GB | Qwen3 235B-A22B Instruct 2507*, Qwen3.5 122B-A10B*, Mistral Small 4 119B* | Qwen3.5 9B (7.9 GB) |
| 512 GB | Qwen3 235B-A22B Instruct 2507*, Qwen3.5 122B-A10B*, Mistral Small 4 119B* | Qwen3.5 9B (7.9 GB) |
* Not in Twoody's catalogue; the others install in one click.
Why memory bandwidth sets the speed
To write each word, the computer reads the model's active weights from memory. Speed therefore follows memory bandwidth more than processor power. Qwen3 8B, 4-bit, on a few machines:
| Computer | Memory bandwidth | Writes |
|---|---|---|
| M1 | 68 GB/s | 8.7–12 tokens/s |
| M2 | 100 GB/s | 13–19 tokens/s |
| M4 | 120 GB/s | 15–21 tokens/s |
| M5 | 153 GB/s | 19–27 tokens/s |
| M4 Pro | 273 GB/s | 31–45 tokens/s |
| M4 Max | 546 GB/s | 53–76 tokens/s |
| M5 Max | 614 GB/s | 72–104 tokens/s |
| GeForce RTX 4090 (24 GB) | 1,008 GB/s | 100+ tokens/s |
4-bit or 8-bit?
Models are shared compressed. 4-bit files (Q4_K_M, Q4_0 for Gemma, MXFP4 for gpt-oss) are half the size of 8-bit ones for a small loss in quality: it is what Twoody installs (3-bit for the larger Qwen3.6 35B-A3B, so that it fits a 24 GB Mac) and what this calculator counts. 8-bit is a little finer and needs twice the memory.
Tokens per second: what it feels like
A token is about three quarters of a word. You read silently at 5 to 6 tokens per second.
| under 5 | slower than you read |
| 5 to 10 | about your reading speed |
| 10 to 20 | faster than you read |
| 20 to 50 | near instant |
| over 50 | feels instant |
A model that thinks first writes 300 to 3,000 tokens of reasoning before its answer: at 10 tokens per second, from half a minute to five minutes. In Twoody, the Fast, Thinking and Deep settings choose how long it thinks.
Reading a document
Before answering about a document, the model reads it — and that is where recent chips gain the most. Qwen3 8B reading a 20-page document, with 16 GB:
| M1 | 3 min |
| M4 | 76 s |
| M5 | 27 s |
Mac or PC?
A Mac shares its memory between the processor and the graphics: a 64 GB Mac runs models no 24 GB graphics card holds. A PC graphics card has less memory but reads it very fast: with a model that fits, an RTX 4090 writes faster than any Mac. Twoody is in private beta on Mac, Windows and Linux: on Windows, it uses the graphics card through Vulkan when the driver supports it; on Linux, it runs the model on the processor.
The best local LLM, by computer
Method and sources
The memory is not this site's estimate: it is the Twoody app's own calculation, with the same files and the same conversation length — the model's file, the file it reads images with and 0.3 GB to use it, the conversation (in q8_0, 8 bits, for a hybrid model such as Qwen3.5; in 16 bits otherwise), the state of its recurrent layers and 0.3 GB of working memory. Measured with llama.cpp, Qwen3.5 9B with a conversation of 32k tokens takes 7.4 GB, as computed. The models are those of Twoody's catalogue, checked on Hugging Face on October 15, 2026, and a few larger ones it does not offer, counted the same way.
Each machine's speed is computed from its measured speed on llama.cpp's reference model (LLaMA 2 7B, 4-bit), scaled by what each model reads per word, with the memory bandwidths of Apple's and NVIDIA's specifications. Checked against 20 published measurements that were not used to build it: median error 4.5 %, largest 23 %, all within the ranges shown.
Figures reviewed on October 16, 2026.
Sources
- www.apple.com/newsroom/
- support.apple.com/specs
- github.com/ggml-org/llama.cpp/discussions/4167
- github.com/ggml-org/llama.cpp/discussions/15013
- github.com/ggml-org/llama.cpp/discussions/10879
- github.com/ggml-org/llama.cpp/discussions/15396
- www.nvidia.com/en-us/geforce/graphics-cards/compare/
- huggingface.co/models
- www.sciencedirect.com/science/article/abs/pii/S0749596X19300786
Questions about running an LLM locally
Can I run a local LLM with 8 GB of memory?
Yes, a small one. On a Mac with Apple silicon and 8 GB, the strongest models that run comfortably are Qwen3.5 4B, Qwen3 4B Instruct 2507 and Nemotron 3 Nano 4B. All install in one click from Twoody's catalogue. Twoody suggests Qwen3.5 4B by default, which needs 4 GB of memory there: one of the models it has measured with its memory, its tools and your documents. Close heavy apps while you use it.
Which LLM for a Mac with 16 GB?
On a Mac with Apple silicon and 16 GB, the strongest models that run comfortably are Gemma 4 12B, GLM-4.6V Flash and Qwen3.5 9B. All install in one click from Twoody's catalogue. Twoody suggests Qwen3.5 9B by default, which needs 7.4 GB of memory there: one of the models it has measured with its memory, its tools and your documents.
How much memory does a 70B model need?
About 45 GB in Twoody, with a conversation of 16k tokens: a Mac with 64 GB runs Llama 3.3 70B, at 4.2–15 tokens/s depending on the chip. Mixtures of experts such as gpt-oss 120B need more memory (60 GB) but write much faster.
Is a 4-bit model much worse than an 8-bit one?
Slightly. For writing, summaries and questions about documents the difference is hard to notice, and 4-bit halves the memory and doubles the speed: it is what Twoody installs, and what this calculator counts.
Mac or NVIDIA graphics card for a local LLM?
A graphics card writes faster with a model that fits its memory (8 to 32 GB); a Mac holds much larger models in its unified memory (up to 512 GB). For one person working on documents, a Mac with 16 to 64 GB is the simplest choice.
Does a local LLM make a MacBook Air too hot?
A MacBook Air has no fan: after a few minutes of continuous writing it slows down to stay cool. For answers of normal length it is not noticeable; for long sessions, a MacBook Pro or a Mac mini keeps its speed.
Do I need to install anything besides Twoody?
No: Twoody downloads the model and runs it with its own engine, llama.cpp, in one click. On a Mac with Apple silicon, it can also install MLX, Apple's engine, if you ask it to (Models › “Advanced”).
How accurate are these figures?
The memory is what the Twoody app shows, computed the same way: for Qwen3.5 9B with a conversation of 32k tokens, llama.cpp measured 7.4 GB, as computed. The speeds are estimates: on 20 published measurements that were not used to build the formula, the median error is 4.5 % and the largest 23 %; each fell within the range shown. Your speed also depends on the app, its version and what else the computer is doing.
Try Twoody before it opens to everyone.
The private beta is free, on Mac, Windows and Linux, and on iPhone and Android with Twoody on your computer. Leave your email to ask for an invitation: we will write to you when it is ready for you.