# Can you run Qwen3.8-27B locally? Yes, on a graphics card you might already own

> Vlad Cealicu on running Alibaba's Qwen3.8-27B at home. The 4-bit file is 16 to 19 GB, so it fits a 24GB graphics card or a 24GB Mac, it writes about 48 to 70 tokens a second on an RTX 4090 or 5090, and why you'd run it for privacy and control rather than to save money on the API.

- Canonical: https://www.anchorterminal.com/blog/run-qwen3-8-27b-locally
- Markdown: https://www.anchorterminal.com/blog/run-qwen3-8-27b-locally.md (~2,400 tokens)
- Slim: https://www.anchorterminal.com/blog/run-qwen3-8-27b-locally.min.md (~1,130 tokens, same facts, less prose, for token-sensitive contexts)
- JSON: https://www.anchorterminal.com/blog/run-qwen3-8-27b-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 28 October 2026 and may change before then. It isn't on the blog yet.

By Vlad Cealicu, 28 October 2026. 765 words, about 3 min to read.

This is the good news post. After [Muse Spark Open](/blog/run-muse-spark-open-locally) and [Kimi K3](/blog/run-kimi-k3-locally), which need a rack, and [GLM-5.3-Flash](/blog/run-glm-5-3-flash-locally), which needs one big machine, Qwen3.8-27B is the open model a lot of people can run on what they already have.

It's also the most downloaded open model released since June. Alibaba released it on 14 August under Apache 2.0, and its repository has 17,870 likes and 16.4 million downloads on Hugging Face, 7.8 million of them in the last 30 days, with another 19.4 million for Unsloth's files [[1]](#ref-1) [[2]](#ref-2).

## How big is Qwen3.8-27B?

27 billion parameters, all of them used for every token, so it's dense rather than a mixture of experts. It has a context of 262,144 tokens that Alibaba says extends to a million, it reads images and video, and it has a multi-token prediction head that can speed it up [[1]](#ref-1). The BF16 weights are 55.6 GB, FP8 30.9 GB and NVIDIA's NVFP4 21.9 GB [[1]](#ref-1) [[3]](#ref-3).

- 55.6 GB: the full BF16 weights
- 16.5 GB: Unsloth's 4-bit file
- 9.8 GB: Unsloth's 2-bit file

Qwen3.8-27B's files, 17 October 2026. Sources [1] [2].

## What hardware runs it?

At 4-bit, anything with 24 GB of memory for the model. A 24GB graphics card like an RTX 3090, 4090 or 5090, a Mac with 24GB, a Strix Halo mini PC or a DGX Spark [[2]](#ref-2). Unsloth says the 4-bit file runs on a 16GB RTX 5080 too, with less room for context [[2]](#ref-2).

## How fast is it?

Fast enough that you're the slow part. These are the testers' own numbers, not ours.

On an RTX 5090 one Hugging Face user measured 70.5 tokens a second with llama.cpp, and with the model's multi-token prediction turned on, about 104 on code and 88 on prose. Federico Viticci got 59 in a single run with LM Studio for MacStories [[4]](#ref-4) [[5]](#ref-5). An RTX 4090 does about 48 [[6]](#ref-6). The 256GB M5 Ultra Mac Studio does 48 in the same MacStories test as the 5090 [[4]](#ref-4). A 128GB Strix Halo does about 17 with multi-token prediction, on a 5-bit file at 64,000 tokens of context, and AMD says up to 24.5 [[7]](#ref-7) [[8]](#ref-8). A DGX Spark does 11.4, or 24.7 with multi-token prediction [[9]](#ref-9). A 24GB M6 Mac mini does 9.8, or 17.5 with it, with room for about 30,000 tokens of context [[10]](#ref-10).

Qwen3.8-27B, tokens a second for one user, different builds and settings

| Setup | Tokens a second |
| --- | --- |
| RTX 5090, multi-token prediction, code | 104 |
| RTX 5090, multi-token prediction, prose | 88 |
| RTX 5090 | 71 |
| RTX 4090 | 48 |
| 256GB Mac Studio, M5 Ultra | 48 |
| DGX Spark, multi-token prediction | 25 |
| 24GB Mac mini, multi-token prediction | 18 |
| 128GB Strix Halo, 5-bit, 64K context, multi-token prediction | 17 |

Each series has a zero-based scale.

## Is it cheaper than the API?

Not really. The cheapest provider on OpenRouter charges $0.50 per million output tokens, and Alibaba charges $2.55 [[11]](#ref-11). The cheapest used RTX 3090s on eBay are about $1,374 to $1,395 [[12]](#ref-12). That's about 2.8 billion output tokens at the cheapest price, which a 3090 writing 45 to 70 tokens a second would take about one and a quarter to two years of non-stop use to get through [[13]](#ref-13). And if the whole PC draws about 400 watts, the electricity alone is about a third of a dollar per million tokens at $0.17 a kWh, close to what the cheapest API charges [[13]](#ref-13).

New cards are worse. The cheapest RTX 5090s on eBay averaged $6,231 in the first half of October and were $6,876 by the 17th, against $1,999 at launch [[14]](#ref-14).

- $1,395: a used RTX 3090
- 2.8 billion: output tokens for the same money
- $0.33: of electricity per million tokens, at 400 watts and $0.17 a kWh

A used RTX 3090 against the cheapest API, at $0.50 per million output tokens. Our arithmetic from [11] [12] [13].

So you'd run it for privacy and control, not to save money. Your data stays on your machine, it works when the internet doesn't, and nobody can change the model under you.

## What I'd do

Run it if you already have the hardware. If you have a 24GB card or a Mac with 24GB or more, this is the open model to try first, and Apache 2.0 means you can build on it without asking anyone. Don't buy a card to save money on the API. It's not Kimi K3 or GLM-5.3-Flash, and Alibaba's own table puts it ahead of Qwen3.7-Plus on most agent tasks, which I'd want to check on my own work before believing [[1]](#ref-1).

That's the series for now, from a rack to a desk. If there's a model you'd like me to do next, tell me.

Qwen3.8-27B now has [its own listing](/tools/qwen3-8-27b) graded by our open-weight checklist, with its builds, the memory each needs and the command that starts it, and [a page comparing it with Gemma 4 31B](/compare/gemma-4-31b-it-vs-qwen3-8-27b), the other dense model of about the same size. [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] Alibaba Qwen, Qwen3.8-27B on Hugging Face. Source for 27B dense parameters, the 262,144-token context extendable to a million, image and video input, the multi-token prediction head, the Apache 2.0 licence, the BF16 and FP8 sizes, the comparison table, and the likes and the all-time and 30-day downloads. https://huggingface.co/Qwen/Qwen3.8-27B

[2] Unsloth, Qwen3.8 guide and unsloth/Qwen3.8-27B-GGUF. Source for the file sizes, the memory each needs, the 4-bit file running on an RTX 5080 and the GGUF downloads. https://unsloth.ai/docs/models/qwen3.8

[3] NVIDIA, Qwen3.8-27B-NVFP4 on Hugging Face. Source for the 21.9 GB NVFP4 build. https://huggingface.co/nvidia/Qwen3.8-27B-NVFP4

[4] Federico Viticci, M5 Ultra Mac Studio review, MacStories, 21 September 2026. Source for 59 tokens a second on an RTX 5090 with LM Studio on Windows and 48 on the M5 Ultra with oMLX on the same 6K-token task, single runs. https://www.macstories.net/stories/m5-ultra-mac-studio-review-the-dream-mac-for-local-ai-agents/

[5] amerm137 and SlavikF, unsloth/Qwen3.8-27B-GGUF discussion 14, August 2026. Source for 70.5 tokens a second on an RTX 5090 without multi-token prediction, and 103.7 to 106.2 on code and 87 to 89 on prose with it, from one tester. Users' reports. https://huggingface.co/unsloth/Qwen3.8-27B-GGUF/discussions/14

[6] erdal, unsloth/Qwen3.8-27B-GGUF discussion 11, 14 August 2026. Source for about 48 tokens a second on an RTX 4090. A user's report. https://huggingface.co/unsloth/Qwen3.8-27B-GGUF/discussions/11

[7] erstmalreden, unsloth/Qwen3.8-27B-GGUF discussion 41, 15 August 2026. Source for 16.68 tokens a second on a 128GB Strix Halo with a 5-bit file at 64K context and multi-token prediction, on Windows. A user's report. https://huggingface.co/unsloth/Qwen3.8-27B-GGUF/discussions/41

[8] AMD, run Qwen3.8-27B on Ryzen AI Max and Radeon, 14 August 2026. Source for "up to 24.5" tokens a second on a Ryzen AI Max+ 395 and the release date. https://www.amd.com/en/blogs/2026/run-qwen-3-8-27b-on-amd-ryzen-ai-max-and-radeon-graphics-cards-day-0.html

[9] helge, NVIDIA developer forums, 15 August 2026. Source for 11.4 tokens a second on a DGX Spark without multi-token prediction and 24.7 with it. A user's report. https://forums.developer.nvidia.com/t/qwen3-8-27b-nvfp4-on-a-single-dgx-spark-up-to-1m-context-vllm-mtp-measurements/380244

[10] lpalbou, mlx-community/Qwen3.8-27B-4bit discussion 4, 30 September 2026. Source for 9.8 and 17.5 tokens a second on a 24GB M6 Mac mini and about 30K of usable context. A user's report. https://huggingface.co/mlx-community/Qwen3.8-27B-4bit/discussions/4

[11] OpenRouter, Qwen3.8-27B endpoints. Source for $0.50 per million output tokens from Tenstorrent and $2.55 from Alibaba. https://openrouter.ai/api/v1/models/qwen/qwen3.8-27b/endpoints

[12] GPU Poet, RTX 3090 prices for October 2026. Source for the cheapest used RTX 3090s on eBay averaging $1,374 in October and $1,395 on 17 October. https://gpupoet.com/gpu/learn/price/october-2026/nvidia-geforce-rtx-3090

[13] SergeySS8, unsloth/Qwen3.8-27B-GGUF discussion 68. Source for 45 to 70 tokens a second on an RTX 3090 with multi-token prediction. A user's report. The 2.8 billion tokens, the time to use them and the electricity, at an assumed 400 watts for the whole PC and $0.17 a kWh, are our arithmetic. https://huggingface.co/unsloth/Qwen3.8-27B-GGUF/discussions/68

[14] GPU Poet, RTX 5090 prices for October 2026. Source for the cheapest eBay listings averaging $6,231 over 1 to 17 October, $6,876 on 17 October, and the $1,999 launch price. https://gpupoet.com/gpu/learn/price/october-2026/nvidia-geforce-rtx-5090

---

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