# Can you run GLM-5.3-Flash locally? Yes, on one big machine

> Vlad Cealicu on what it takes to run Z.ai's GLM-5.3-Flash, 320 billion parameters with 18 billion active, at home. One 256GB Mac Studio runs a mostly 4-bit build at about 30 tokens a second, a 128GB Strix Halo mini PC runs a squeezed one at about 29, two DGX Sparks or two used RTX 3090s with plenty of RAM get into the 20s, and what each costs against an API at 50 cents per million tokens.

- Canonical: https://www.anchorterminal.com/blog/run-glm-5-3-flash-locally
- Markdown: https://www.anchorterminal.com/blog/run-glm-5-3-flash-locally.md (~2,750 tokens)
- Slim: https://www.anchorterminal.com/blog/run-glm-5-3-flash-locally.min.md (~1,430 tokens, same facts, less prose, for token-sensitive contexts)
- JSON: https://www.anchorterminal.com/blog/run-glm-5-3-flash-locally.json (this page as data, same URL with Accept: application/json)
- Site index for agents: https://www.anchorterminal.com/llms.txt (full text: https://www.anchorterminal.com/llms-full.txt)
- API: https://www.anchorterminal.com/api/v1/index.json
- MCP: https://www.anchorterminal.com/mcp (the whole catalogue as tools, no key)
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)
- Updated: 2026-10-21

> **Preview.** This post is scheduled for 27 October 2026 and may change before then. It isn't on the blog yet.

By Vlad Cealicu, 27 October 2026. 903 words, about 4 min to read.

GLM-5.3-Flash is the first big open model in this series that a person can run at home at a usable speed, on one machine. That's a nice change after [Muse Spark Open](/blog/run-muse-spark-open-locally) and [Kimi K3](/blog/run-kimi-k3-locally), where the honest answer was a rack.

Z.ai released it on 26 August under the MIT licence, and it's been downloaded 9.2 million times from Hugging Face, 6.57 million of them in the last 30 days, plus 1.55 million for Unsloth's files [[1]](#ref-1) [[2]](#ref-2).

## How big is GLM-5.3-Flash?

320 billion parameters, with 18 billion active for each token. It's a mixture of experts with 288 experts and 8 picked per token, a context of a million tokens, and it reads video, images, files and text [[1]](#ref-1) [[3]](#ref-3). The official weights are 328 GB in FP8 and 643 GB in BF16, and NVIDIA publishes a 177.5 GB build it calls FP4-Local, which it validated on two DGX Sparks or two RTX PRO 6000s [[1]](#ref-1) [[4]](#ref-4).

The 18 billion active is what makes it work at home. Each token only reads about 11 GB of a 4-bit file, against about 58 GB for Kimi K3, so a machine with fast memory can keep up [[5]](#ref-5).

- 328 GB: the official FP8 weights
- 200 GB: Unsloth's 4-bit file
- 93 GB: Unsloth's smallest 1-bit file

GLM-5.3-Flash's weights, 17 October 2026. Sources [1] [2] [4].

## What do the smaller files cost you?

Unsloth publishes files from 93 GB at 1-bit to 341 GB at 8-bit, and says how often each picks the same top token as the full model. 71 per cent at 1-bit, 78 at 2-bit, 82 to 86 at 3-bit and 92 at 4-bit [[2]](#ref-2). NVIDIA reports 91.3 per cent on GPQA Diamond for its FP4-Local build, against 92.1 for the full BF16 weights [[4]](#ref-4).

How often each file agrees with the full model's top token, per cent

| File | Per cent |
| --- | --- |
| 4-bit | 92 |
| 3-bit, the bigger file | 86 |
| 2-bit | 78 |
| 1-bit | 71 |

Each series has a zero-based scale.

## What hardware runs it, and how fast?

This is where GLM-5.3-Flash is different. People have run it on nearly every kind of machine I looked at for Muse, and published the numbers. They're the testers' own numbers, not ours, and the setups differ, so read them as what's possible rather than a fair race.

A 256GB Mac Studio with the M5 Ultra runs a mixed 4 and 8-bit MLX build at 28 to 32 tokens a second on short answers as the prompt grows from 3,400 to 61,000 tokens, and 41 on code, in Federico Viticci's review at MacStories [[6]](#ref-6). A 128GB Strix Halo mini PC runs the tester's own 2 to 3-bit build at about 29 tokens a second with multi-token prediction, with Vulkan beating ROCm, in a Hugging Face user's test [[7]](#ref-7). Two DGX Sparks run a third-party 4-bit build at 20 to 23 tokens a second with multi-token prediction and a patched vLLM [[8]](#ref-8). Two used RTX 3090s with 128GB of DDR4 reach 24 on a forked llama.cpp [[16]](#ref-16). A CPU-only server with two Xeons and 1TB of DDR5 runs the 8-bit file at 20 to 21 tokens a second on a heavily patched llama.cpp, against 6.9 on the stock one, and that 1TB of DDR5 alone is $36,000 to $44,500 at today's prices [[9]](#ref-9) [[15]](#ref-15). A single RTX 5090 with 64GB of RAM manages about 6, reading the experts from SSD as it goes, with patches of its own [[10]](#ref-10).

GLM-5.3-Flash, tokens a second for one user, different builds and settings

| Setup | Tokens a second |
| --- | --- |
| 256GB Mac Studio, mixed 4 and 8-bit, short answers | 31 |
| 128GB Strix Halo, 2 to 3-bit, with multi-token prediction | 29 |
| Two RTX 3090s and 128GB DDR4, forked llama.cpp | 24 |
| Two DGX Sparks, 4-bit, with multi-token prediction | 23 |
| Two Xeons, 1TB DDR5, 8-bit | 21 |
| One RTX 5090, streaming from SSD | 6 |

Each series has a zero-based scale.

## What does that cost?

A 256GB Mac Studio is $9,499 or $10,799, because Apple's own pages disagree on whether 256GB needs the bigger chip [[11]](#ref-11). A Minisforum MS-S1 Max with 128GB is £3,519 in the UK and $3,879 in the US, the cheapest way into this model, if you accept a 2 to 3-bit build [[12]](#ref-12). Two DGX Sparks are $13,900 [[13]](#ref-13). When I checked on 17 October, the 512GB Mac Studio, which would hold the full FP8 weights, still wasn't on sale and had no price [[11]](#ref-11).

The API is cheap. Z.ai charges $0.15 per million tokens in and $0.50 per million out, and DeepInfra $0.075 and $0.25 [[14]](#ref-14). The Mac Studio buys about 19 billion output tokens from Z.ai, and the MS-S1 Max about 7.8 billion, which at 30 tokens a second is about 20 years and 8 years of non-stop writing [[14]](#ref-14).

- 19 billion: output tokens for the price of a 256GB Mac Studio
- 7.8 billion: for the price of a 128GB MS-S1 Max

What the hardware would buy from Z.ai's API, at $0.50 per million output tokens. Our arithmetic from [11] [12] [14].

## What I'd do

If you want it for the price, use the API. If your data shouldn't leave the building, or you want it to work when the internet doesn't, GLM-5.3-Flash is the first big model in this series I'd say yes to. The 256GB Mac Studio is the simplest, it runs a mostly 4-bit build fast enough to use, and the MS-S1 Max is the cheapest, if a squeezed build is good enough for what you do. I'd test the 3-bit build on my own work before trusting the 2-bit one.

The last post in the series is the small one, Qwen3.8-27B, which runs on a graphics card a lot of people already own.

GLM-5.3-Flash now has [its own listing](/tools/glm-5-3-flash) graded by our open-weight checklist, with its builds, the memory each needs and the command that starts it, next to [GLM-5.3](/tools/glm-5-3), and [a page comparing the two](/compare/glm-5-3-vs-glm-5-3-flash). [Your cousins](/blog/your-cousins-open-models-for-agents) is about which open models an agent should pick.

## References

Read on 17 October 2026.

[1] Z.ai, GLM-5.3-Flash on Hugging Face, with its config and the BF16 repository. Source for 320B total and 18B active parameters, 288 experts with 8 per token, the million-token context, the MIT licence, the FP8 and BF16 sizes, the upload on 26 August and the downloads, all-time and 30-day. https://huggingface.co/zai-org/GLM-5.3-Flash

[2] Unsloth, GLM-5.3-Flash guide and unsloth/GLM-5.3-Flash-GGUF. Source for the file sizes, the memory each needs, the top-token agreement by bits (from Unsloth's table) and the GGUF downloads, all-time and 30-day. https://unsloth.ai/docs/models/glm-5.3-flash

[3] Z.ai, GLM-5.3-Flash developer guide. Source for video, image, file and text input. https://docs.z.ai/guides/llm/glm-5.3-flash

[4] NVIDIA, GLM-5.3-Flash-FP4-Local on Hugging Face. Source for the 177.5 GB build, its validation on two DGX Sparks or two RTX PRO 6000s and GPQA Diamond 91.3 against 92.1 per cent. https://huggingface.co/nvidia/GLM-5.3-Flash-FP4-Local

[5] Anchor Terminal, can you run Kimi K3 locally, 25 October 2026. The 11 GB and 58 GB a token are our arithmetic from active parameters and file sizes, as are the token totals below. https://www.anchorterminal.com/blog/run-kimi-k3-locally

[6] Federico Viticci, M5 Ultra Mac Studio review, MacStories, 21 September 2026. Source for 31, 32 and 28 tokens a second at 3.4K, 13.6K and 61K prompts on short answers and 41 on code, with a mixed 4 and 8-bit MLX build. https://www.macstories.net/stories/m5-ultra-mac-studio-review-the-dream-mac-for-local-ai-agents/

[7] j2ipper, unsloth/GLM-5.3-Flash-GGUF discussion 27, 14 October 2026. Source for about 29 tokens a second on a 128GB Strix Halo with the poster's own 2 to 3-bit build and multi-token prediction, and Vulkan at 29.6 against ROCm at 21.5. A user's report. https://huggingface.co/unsloth/GLM-5.3-Flash-GGUF/discussions/27

[8] H-K-B, GLM-5.3-Flash discussion 52, 15 September 2026. Source for 20 to 23 tokens a second for one user on two DGX Sparks with a third-party NVFP4 build, multi-token prediction and a patched vLLM. A user's report. https://huggingface.co/zai-org/GLM-5.3-Flash/discussions/52

[9] goldvet, GLM-5.3-Flash discussion 60, 9 October 2026. Source for 20 to 21 tokens a second on two Xeon Gold 6530s with 1TB of DDR5 on a heavily patched llama.cpp, and 6.9 on the stock one. A user's report. https://huggingface.co/zai-org/GLM-5.3-Flash/discussions/60

[10] xMytze, unsloth/GLM-5.3-Flash-GGUF discussion 22, 24 September 2026. Source for about 6 tokens a second on one RTX 5090 with 64GB of RAM, streaming experts from SSD with its own patches. A user's report. https://huggingface.co/unsloth/GLM-5.3-Flash-GGUF/discussions/22

[11] Daring Fireball, configurations and pricing for the new Mac Studios, 25 August 2026, and Apple's store page. Source for $5,499 plus $4,000 for 256GB, Apple's specs page tying 256GB to the 36-core chip, which would make it $10,799, and the 512GB option "coming late October". https://daringfireball.net/2026/08/configurations_and_pricing_for_new_mac_minis_and_mac_studios

[12] Minisforum, MS-S1 Max, UK and US stores, read 17 October 2026. Source for £3,519 and $3,879 for the 128GB model. https://www.minisforum.uk/products/minisforum-ms-s1-max

[13] The Register, 2 October 2026. Source for the 128GB DGX Spark at $6,950. https://www.theregister.com/systems/2026/10/02/nvidia-debuts-4999-dgx-spark-with-half-the-ram-and-storage-amid-memory-crunch/5300622

[14] OpenRouter, GLM-5.3-Flash endpoints. Source for $0.15 and $0.50 per million tokens from Z.ai and $0.075 and $0.25 from DeepInfra. https://openrouter.ai/api/v1/models/z-ai/glm-5.3-flash/endpoints

[15] DatacenterDisk, DDR5 server RAM prices, updated 17 October 2026. Source for $35.08 a GB for the cheapest 64GB sticks and $43.46 at the median, from which the 1TB totals are our arithmetic. https://datacenterdisk.com/server-ram/ddr5

[16] neuralll, unsloth/GLM-5.3-Flash-GGUF discussion 23, 25 September 2026. Source for 24 tokens a second on two RTX 3090s with 128GB of DDR4 on a forked llama.cpp with a GPU expert cache, against 12 on the stock one. A user's report. https://huggingface.co/unsloth/GLM-5.3-Flash-GGUF/discussions/23

---

Query this catalogue directly over MCP: https://www.anchorterminal.com/mcp (Streamable HTTP, no key). How: https://www.anchorterminal.com/agents/#mcp