AI News

Best Local AI Models for 128GB: Strix Halo PCs and 128GB Macs

Short answer: with 128 GB, run gpt-oss-120b. It has the highest score on our board (98.0, likely 96.0 to 100.0; Qwen3.8 27B’s 2-bit file is within range of it), tied for the top Kingy Lab score (98%), fits with 23.8 GB to spare, and wrote 33.8 tokens a second on real jobs on our 128GB Lucy AI Studio. For quicker replies, Qwen3.6 35B A3B tied with it on Kingy Lab (98%), scored 89.7 on our board (the ranges just overlap, so the board can’t separate the two), and was nearly twice as fast (62.0 tokens a second).

Lucy AI Studio sent free by Cephalon, which paid for an earlier Kingy AI video. Cephalon has no say in results. Mac and laptop bought by Kingy AI.

Who this page is for

People with 128 GB of shared memory: an AMD Ryzen AI Max+ 395 (“Strix Halo”) mini PC like our Lucy AI Studio, or a 128GB Mac. By our fit rule both can use 96.0 GB. All speeds on this page are from the Lucy; we haven’t measured a 128GB Mac.

Top picks for 128GB

Job Our pick File to download Score Speed
Coding and building apps gpt-oss-120b gpt-oss-120b-F16.gguf (65.37 GB) coding jobs 99.2; board 98.0 (likely 96.0 to 100.0) 33.8 tok/s on real jobs
Lucy AI Studio, Ryzen AI Max+ 395, 128 GB, Kingy measured Nov 14-15, 2026
Writing to instructions Qwen3.6 35B A3B or gpt-oss-120b Ollama qwen3.6:35b (Q4_K_M) (21.72 GB) Kingy Lab writing-to-rules: Qwen3.6 100%, gpt-oss-120b 99% 62.0 and 33.8 tok/s on real jobs
Lucy AI Studio, Ryzen AI Max+ 395, 128 GB, Kingy measured Nov 14-15, 2026
Everyday questions and chat Qwen3.6 35B A3B Ollama qwen3.6:35b (Q4_K_M) (21.72 GB) Kingy Lab everyday 99% (gpt-oss-120b: 99%) 62.0 tok/s on real jobs
Lucy AI Studio, Ryzen AI Max+ 395, 128 GB, Kingy measured Nov 14-15, 2026
Long documents and files Qwen3.6 35B A3B or gpt-oss-120b Ollama qwen3.6:35b (Q4_K_M) (21.72 GB) long-document and file jobs: Qwen3.6 100.0, gpt-oss-120b 97.6 (ranges overlap) 62.0 and 33.8 tok/s on real jobs
Lucy AI Studio, Ryzen AI Max+ 395, 128 GB, Kingy measured Nov 14-15, 2026

gpt-oss-120b has 117 billion parameters but uses 5.1 billion for each token (OpenAI’s model card), which is why it’s quicker than the 27B and 31B dense models on this machine. The file is a GGUF copy that keeps OpenAI’s own MXFP4 expert weights (OpenAI trained them that way), so the 65.37 GB download isn’t a further-squeezed copy. Qwen3.8 27B’s 2-bit file scored nearly as high on the board (97.9) but wrote only 18.8 tokens a second here; at 128 GB there’s no reason to squeeze it.

Model File we scored Memory needed, 8K context 128GB machine (96.0 GB usable) Board score (on our Lucy)
Qwen3 8B Qwen3-8B-Q4_K_M.gguf (5.03 GB) 6.86 GB Fits, 89.1 GB spare 74.5
Qwen3.8 27B, 2-bit file Qwen3.8-27B-UD-IQ2_S.gguf (8.37 GB) 9.80 GB Fits, 86.2 GB spare 97.9
Qwen3.8 27B, 3-bit file Qwen3.8-27B-UD-Q3_K_XL.gguf (13.15 GB) 15.05 GB Fits, 81.0 GB spare 93.1
Gemma 4 26B A4B gemma-4-26B-A4B-it-Q4_K_M.gguf (16.80 GB) 18.89 GB Fits, 77.1 GB spare 87.2
Gemma 4 31B gemma-4-31B-it-Q4_K_M.gguf (18.69 GB) 22.22 GB Fits, 73.8 GB spare 91.2
Qwen3.6 35B A3B Ollama qwen3.6:35b (Q4_K_M) (21.72 GB) 24.08 GB Fits, 71.9 GB spare 89.7
gpt-oss-120b gpt-oss-120b-F16.gguf (65.37 GB) 72.24 GB Fits, 23.8 GB spare 98.0

How fast they run on our 128GB Lucy

Real-job speed is how fast each model wrote during the board: tokens written divided by the time spent writing (llama.cpp’s own timings, prompt reading left out), over the 27 board runs per model, November 14 to 15, 2026 (UTC).

Bar chart of real-job speed on the 128GB Lucy: Qwen3.6 35B 62.0 tok/s, gpt-oss-120b 33.8, Gemma 4 31B 10.5
Model (file as listed) Real-job speed Board score
Qwen3.6 35B A3B 62.0 tok/s 89.7
Gemma 4 26B A4B 54.5 tok/s 87.2
Qwen3 8B 36.5 tok/s 74.5
gpt-oss-120b 33.8 tok/s 98.0
Qwen3.8 27B, 2-bit file 18.8 tok/s 97.9
Qwen3.8 27B, 3-bit file 14.2 tok/s 93.1
Gemma 4 31B 10.5 tok/s 91.2

We also timed a few models we haven’t put on the board, with llama.cpp’s speed tool and 8,192 tokens already loaded (November 12, 2026): LFM2.5 24B A2B 87.8 tokens a second, Ornith 1.5 35B A3B 58.9, and Gemma 4 26B A4B 44.8 in a larger 5-bit file. Without board scores we can’t rank them for quality.

How to run it

The easy way: LM Studio. It has a built-in model downloader. Open the Discover tab (Cmd+2 on a Mac, Ctrl+2 on Windows or Linux), paste the repository name shown for each model below into the search bar, and pick the file with the same quant name (LM Studio’s guide).

The direct way: llama.cpp, the engine we test with. Install it with winget install llama.cpp on Windows or brew install llama.cpp on a Mac (install guide). -hf downloads a model from Hugging Face and -c 8192 sets the 8K context our fit numbers assume. -hf also downloads and loads a model’s vision file if it has one (about 1.2 GB for Gemma 4); our fit numbers are for text only, so the commands below add --no-mmproj to skip it (server options).

  • gpt-oss-120b: llama-server -hf unsloth/gpt-oss-120b-GGUF:F16 -c 8192 --no-mmproj, or search unsloth/gpt-oss-120b-GGUF in LM Studio and pick F16.
  • Qwen3.6 35B A3B: the file we scored is Ollama’s. Install Ollama, then run ollama run qwen3.6:35b. That tag is the same download as qwen3.6:latest on Ollama’s tag list (checked November 18, 2026).
  • Gemma 4 31B: llama-server -hf lmstudio-community/gemma-4-31B-it-GGUF:Q4_K_M -c 8192 --no-mmproj, or search lmstudio-community/gemma-4-31B-it-GGUF in LM Studio and pick Q4_K_M.

MLX on a Mac. Apple’s own MLX-LM runs models converted to MLX format: pip install mlx-lm, then mlx_lm.generate --model <an MLX model on Hugging Face> --prompt "hello" (or mlx_lm.chat for a chat session). MLX files are different downloads from the ones above, with different sizes, and we didn’t score or time them, so check the size before you download.

How we tested

Kingy tested every score on this page on our own machines. Nothing here is copied from a model maker’s chart.

  • Board score. Kingy’s local AI board v0.2 gives each model the same 9 real jobs: a landing page, a budget tracker app, a browser game, a Python bug fix, a chart from a messy spreadsheet, questions about a long document, a job that means searching 37 files with tools, a motion title and a video cut. Scripts check each result; no AI judges another AI. Every job runs 3 times, with the makers’ own settings and the exact files named on this page (checked by fingerprint). The board score is the average of 7 of those jobs, 21 runs in all. The landing page and the motion title are run but not counted, because they didn’t tell the models apart: every model on our Lucy scored 90.2 or more on the landing page, and motion title scores swung too much from one run to the next. The main runs were on our Lucy AI Studio (AMD Ryzen AI Max+ 395, 128 GB), November 14 to 15, 2026 (UTC). If a check script stopped early on a run, the checks it didn’t reach count as fails. “Likely range” is a rough 95% range around the average: when two ranges overlap, we treat the board as unable to tell those two models apart.
  • Writing and everyday scores come from Kingy Lab’s 120-task test (November 7 to 12, 2026), which marks 60 everyday jobs, 30 coding tasks and 30 writing tasks with rules to follow.
  • Fit uses the same rule as our Local AI checker: memory needed = file size + room for 8,192 tokens (8K) of conversation (the “KV cache”, from each model’s official settings file) + 10%. We count a graphics card’s memory minus 1.5 GB for the screen and driver, and, as a cautious rule of our own, two-thirds of a Mac’s memory up to 36 GB (three-quarters above that), leaving the rest for macOS and your apps. macOS itself lets the graphics side use more (about 19 GB of our 24GB MacBook Pro), so a Mac may fit a little more than we say. For Qwen3.6 and Qwen3.8 we leave out a small fixed memory state (under 0.2 GB). GB means a billion bytes. File sizes come from Hugging Face and Ollama, checked November 18, 2026. A “No” here means it doesn’t fit in that memory by this rule.
  • Speed is only shown where we measured it, with the chip and date. Everything else says “not measured yet”. We never estimate a speed on these pages. Speed on real jobs is how fast the model wrote during the 27 board runs: tokens written divided by the time spent writing, from llama.cpp’s own timings, counting every reply (including replies from runs we repeated after a dropped connection). It leaves out the time spent reading the prompt.

What we don’t know yet

  • Speeds on a 128GB Mac. Not measured yet; Apple’s chips have different memory speeds from the Ryzen AI Max+ 395.
  • Bigger models. Models over about 90 GB (for example DeepSeek-class releases) aren’t on the board yet, so this page doesn’t rank them.
  • Image generation at 128 GB. Not measured, so no pick.

Not sure how much memory you have, or using a different chip? The Kingy Local AI checker asks for your machine and shows which of 50 popular models fit, with our measured speeds where we have them.

Sources

  • Kingy local AI board v0.2 and Kingy speed tests, checked 2026-11-18: Kingy’s board v0.2 scores and speed tests behind this page (method above).
  • Kingy Lab: Local AI Models Tested, checked 2026-11-18: Kingy Lab’s 120-task test of 8 open models on the Lucy, November 7 to 12, 2026 (writing, everyday and coding scores).
  • Kingy Local AI checker, checked 2026-11-18: The fit rule used on this page (Kingy verdict standard v0.1).
  • huggingface.co, checked 2026-11-18: Qwen’s Qwen3 8B files and sizes on Hugging Face.
  • huggingface.co, checked 2026-11-18: Unsloth’s Qwen3.8 27B files at the revision we scored (August 20, 2026).
  • huggingface.co, checked 2026-11-18: Unsloth’s Qwen3.8 27B file history: many files, including both we scored, replaced November 10 to 12.
  • huggingface.co, checked 2026-11-18: Gemma 4 26B A4B files and sizes (LM Studio Community).
  • huggingface.co, checked 2026-11-18: Gemma 4 31B files and sizes (LM Studio Community).
  • ollama.com, checked 2026-11-18: Ollama’s Qwen3.6 tags; the 35b tag is the file we scored.
  • huggingface.co, checked 2026-11-18: Qwen’s Qwen3.6 35B A3B model card (total and active parameters).
  • huggingface.co, checked 2026-11-18: gpt-oss-120b file and size (Unsloth).
  • huggingface.co, checked 2026-11-18: OpenAI’s gpt-oss-120b model card (total and active parameters).
  • lmstudio.ai, checked 2026-11-18: LM Studio’s guide to downloading a model.
  • github.com, checked 2026-11-18: llama.cpp install guide.
  • github.com, checked 2026-11-18: llama.cpp server options (-hf, -m, -c).
  • github.com, checked 2026-11-18: Apple’s MLX-LM project.
  • ollama.com, checked 2026-11-18: Ollama download page.

Revision log

  • November 18, 2026: First published. Picks from Kingy board v0.2 (November 14 to 15), Kingy Lab (November 7 to 12) and Lucy speed tests (November 12 to 15).