98% of ChatGPT level
writes text
· needs ~20 GB · Free to download · License: Apache 2.0
Your computer: not set · Set it
This computer has 128 GB of unified memory. A model is listed when it needs no more than 96 GB, the share AMD lets you dedicate to the graphics chip.
The store name is Minisforum MS-S1 MAX. It uses an AMD Ryzen AI Max+ 395 with Radeon 8060S graphics, 40 graphics cores at up to 2900 MHz, and 128 GB of LPDDR5x-8000 on a 256-bit bus. The memory is unified, so the graphics chip draws from the same pool as the rest of the computer. Minisforum ships Windows 11 Pro. AMD lists Ubuntu as a supported system for this processor.
AMD Variable Graphics Memory can dedicate 96 GB of the 128 GB to the Radeon 8060S. That 96 GB block is the pool the fit rule counts. AMD says the graphics chip can address up to 112 GB in total, and that the fast path is to keep the model inside the 96 GB block. The store listing we checked is this 128 GB configuration, not a separate 96 GB computer, and not a model named MS-M1.
The memory bus is 256 GB/s, so a model can fit and still be slow. Minisforum’s figure of 10.87 tokens per second for a 235B Q4 model is for two of these computers clustered together.
Specs checked 16 Oct 2026 on the Minisforum store page and AMD’s note on Variable Graphics Memory.
Slow. You will wait on each reply.
Faster than you can read
It needs about 120 GB of memory, and this computer has about 96 GB free.
98% of ChatGPT level
writes text
· needs ~20 GB · Free to download · License: Apache 2.0
97% of ChatGPT level
writes, codes, and translates
· needs ~18 GB · Free to download · License: Apache 2.0
97% of ChatGPT level
writes text, and quicker than it looks
· needs ~16 GB · Free to download · License: Apache 2.0
95% of ChatGPT level
strong at writing, code, and other languages
· needs ~18 GB · Free to download · License: Apache 2.0
94% of ChatGPT level
fast chat and code, because only a slice wakes up
· needs ~20.4 GB · Free to download · License: Apache 2.0
92% of ChatGPT level
a strong writer, a step behind Gemma 4
· needs ~17 GB · Free to download · License: Gemma Terms of Use
Better for code than an 8B general model, and behind Qwen3.8 27B on the public chat board.
fast help with code
· needs ~19 GB · Free to download · License: Apache 2.0
91% of ChatGPT level
OpenAI's big open model, for huge memory pools
· needs ~65 GB · Free to download · License: Apache 2.0
91% of ChatGPT level
writes text on a 16 GB machine
· needs ~8.2 GB · Free to download · License: Gemma Terms of Use
A step behind the full DeepSeek R1, which scores 1,398 on LMArena, near Claude from late 2024.
talks through the steps before it answers
· needs ~20 GB · Free to download · License: MIT
A step behind Gemma 3 12B, and a clear step ahead of a 3B toy model.
a quick chatbot that fits smaller computers
· needs ~5.2 GB · Free to download · License: Apache 2.0
89% of ChatGPT level
pleasant everyday notes on a laptop
· needs ~5 GB · Free to download · License: Gemma Terms of Use
89% of ChatGPT level
a large chatbot that needs lots of memory
· needs ~40 GB · Free to download · License: Llama 3.3 Community License
89% of ChatGPT level
a solid helper for chat and code
· needs ~14 GB · Free to download · License: Apache 2.0
88% of ChatGPT level
writes text, and it can look at a photo
· needs ~3.4 GB · Free to download · License: Gemma Terms of Use
86% of ChatGPT level
a well-known local coding helper
· needs ~20 GB · Free to download · License: Apache 2.0
86% of ChatGPT level
a multilingual writer
· needs ~20 GB · Free to download · License: CC-BY-NC 4.0
85% of ChatGPT level
careful answers, strong for its size
· needs ~9.1 GB · Free to download · License: MIT
A handy junior for small coding jobs. It is not in the same league as Qwen3.8 27B or a 2026 cloud coding agent.
helps with short bits of code
· needs ~4.7 GB · Free to download · License: Apache 2.0
83% of ChatGPT level
writes and translates across languages
· needs ~5 GB · Free to download · License: CC-BY-NC 4.0
Strong at getting the meaning across, including languages the big chatbots treat lightly. For French, Spanish, or German prose, a large Qwen or Gemma usually sounds more natural.
translates many languages
· needs ~2.5 GB · Free to download · License: CC-BY-NC 4.0
79% of ChatGPT level
short rewrites on a tiny computer
· needs ~2 GB · Free to download · License: Llama 3.2 Community License
Transcription: 93% of Rev.ai
turns speech into text, in the larger file
· needs ~3.1 GB · Free to download · License: MIT
The step up from FLUX.1. In the neighborhood of recent cloud image tools for many prompts, and still behind the best closed models on tricky text and hands.
a large model for still pictures
· needs ~19 GB · Free to download · Non-commercial license
Non-commercial licenseThe weights are free to download. The license is for non-commercial use. FLUX.2 Klein 4B is Apache 2.0.
Transcription: 99% of Rev.ai
turns recordings into text, with less waiting
· needs ~1.6 GB · Free to download · License: MIT
Closer to a performance than Kokoro. Still not a substitute for a directed studio voice.
reads text aloud with more expression
· needs ~2 GB · Free to download · License: MIT
Sharper short clips than the small Wan model, including audio in the LTX stack. It does not replace a cloud video studio or an editor.
makes a short video clip, with sound
· needs ~30 GB · Free to download · License: See Lightricks license
Often compared with DALL·E 3 and Midjourney v6 for prompt following. The license is non-commercial.
makes pictures that follow the prompt
· needs ~12 GB · Free to download · License: Non-commercial FLUX.1 Dev license
The document reader to pick when the scan is more than a clean page of print.
reads text from scans (OCR), including tables
· needs ~5.5 GB · Free to download · License: Apache 2.0
Near the top of small open English voices in public listening tests. Less expressive than a top cloud voice studio.
reads text aloud in a warm voice
· needs ~0.3 GB · Free to download · License: Apache 2.0
Within a small gap of Whisper large-v3 on English, and faster.
turns speech into text, in a smaller file
· needs ~1.5 GB · Free to download · License: MIT
A vision model that reads pages and answers questions about them. More general than a pure OCR engine.
reads text from scans (OCR) and answers about the picture
· needs ~9 GB · Free to download · License: Apache 2.0
The best widely documented open clip model for a 24 GB-and-up machine. Still a short clip, not a scene from a film.
makes a better short video clip
· needs ~26 GB · Free to download · License: Apache 2.0
The best picture model that still feels at home on a 32 GB Mac mini.
makes sharper pictures in a smaller file
· needs ~13 GB · Free to download · License: Apache 2.0
Close to FLUX.1 Dev for a lot of everyday prompts, and well ahead of Stable Diffusion XL.
makes pictures quickly
· needs ~8 GB · Free to download · License: Apache 2.0
A sketch, not a finished ad. Enough to see an idea move.
makes a short, rough video clip
· needs ~8 GB · Free to download · License: Apache 2.0
Transcription: 97% of Rev.ai
turns speech into text, almost anywhere
· needs ~0.5 GB · Free to download · License: MIT
Fine for playful and stylized pictures. A clear step behind FLUX on instructions and realistic scenes.
makes pictures on small computers
· needs ~7 GB · Free to download · License: OpenRAIL++
Invoice OCR: 107% of ABBYY (2022 test)
reads text from scans (OCR), including tilted photos
· needs ~0.2 GB · Free to download · License: Apache 2.0
A fine-tune, not a new architecture. It sits with other 8B chat models, and it is not given a separate ChatGPT percentage here.
A Nous Research fine-tune of Llama 3.1 8B
· needs ~4.7 GB · Free to download · License: Llama 3.1 Community License
Understandable and a bit robotic next to Kokoro.
reads text aloud in a plain voice
· needs ~0.1 GB · Free to download · License: GPL-3.0
Invoice OCR: 84% of ABBYY (2022 test)
reads text from scans (OCR)
· needs under 0.1 GB · Free to download · License: Apache 2.0
A cleanup tool, not a rival to FLUX.
enlarges a picture you already have
· needs ~0.3 GB · Free to download · License: BSD 3-Clause
A standard local search model. Quality of the written answer comes from the chat model beside it.
archives and searches your documents
· needs ~1.2 GB · Free to download · License: MIT
A popular small embedder in Ollama. bge-m3 is the better default when files mix languages.
searches English documents
· needs ~0.3 GB · Free to download · License: Apache 2.0