Kingy AI · Local AI hardware verdicts · Standard v0.1
Can You Run It at Home?
Kingy’s verdict for each model on six common machine types, from a 16GB laptop to a 512GB Mac Studio. Every verdict names the exact version and says how Kingy knows: measured, reported by a named third party, or estimated.
Estimated verdicts are calculated, not run. Kingy works them out from published file sizes and memory bandwidth, and labels them “(estimated)”. A measured or reported result always wins over an estimate.
24 verdicts for 4 models: 1 Kingy measured · 10 Reported · 10 Estimated · 3 No evidence yet. Data updated Oct 18, 2026.
Reference machine: Mac Studio M3 Ultra, 512GB · 384 GB (macOS default; raising iogpu.wired_limit_mb allows more) usable · 819 GB/s. Best verdict first.
Qwen3.8-Flash-Next
Text model · Qwen (Alibaba)
Can you run Qwen3.8-Flash-Next at home? Yes. 512GB Mac Studio · MLX oQ8 (8-bit) with MTP head · oMLX v0.7.0, MTP on · 62.4 tok/s at 8K context · Reported Oct 2, 2026
Yes
Yes. 62.4 tok/s at 8K with MTP on, prefill 845 tok/s (an 8K prompt takes about 10 s), about 208 GB peak. Without MTP, named 4-bit reports show 24 to 31 tok/s on short prompts. Kingy found no Apple runtime for the official FP8 file.
Version
MLX oQ8 (8-bit) with MTP head · oMLX v0.7.0, MTP onTensorFold/Qwen3.8-Flash-Next-MLX-oQ8-MTP (matched by name) · Qwen3.8-Flash-Next-MLX-oQ8-MTP (37 safetensors shards)
Can you run DeepSeek V4.1 Flash 1017 at home? Barely. 512GB Mac Studio · DS4 GGUF, MXFP4 + Q8_0; Engram on SSD · ds4-metal main @ 24e30c0 · 37.6 tok/s at 8K context · Reported Oct 18, 2026
Barely
Barely (reported). ds4-metal runs a native-MXFP4 GGUF at 37.6 tok/s at 8K on an M3 Ultra, but keeps the Engram tables (~40% of the file) on SSD. The 432GB MLX oQ4e needs ~476GB, over the 384GB default, so it fits only by raising iogpu.wired_limit_mb.
Version
DS4 GGUF, MXFP4 + Q8_0; Engram on SSD · ds4-metal main @ 24e30c0 (card requires 3e92959 or later)ivanfioravanti/DeepSeek-V4.1-Flash-DS4-Q4 · DeepSeek-V4.1-Flash-1017-MXFP4.gguf (part1 + part2 joined) @ 4d16135
Can you run MiniMax H3 at home? Barely. 512GB Mac Studio · MLX affine 8-bit, group size 64 · mlx-serve · 20 min per 5.17 s clip at 960x768 (124 frames), 30 steps · Reported Aug 3, 2026
Barely
Barely. Only a reduced 960x768 canvas is reported: a 124-frame clip took about 19.3 min sampling plus 64 s decode at 30 steps with an MLX 8-bit pack. No 1344x768 run is reported for this machine. Reported, not Kingy-tested.
Every verdict follows Kingy’s “Can you run it at home?” standard v0.1. A verdict rates one model on one machine type, names the exact version, and always says how Kingy knows.
Text models
Verdict
What it takes
Yes
Loads entirely in GPU or unified memory, generates at least 20 tokens/s, works at 8K context or more, and starts answering an 8K prompt within 30 seconds. All four must hold.
Yes, but slow
Loads and runs, but generates 5–20 tokens/s, or takes more than 30 seconds to start answering an 8K prompt.
Barely
Runs, but generates under 5 tokens/s, or needs more than 25% of the weights in system RAM (weights read from SSD count too), or fits only at 3-bit or lower, or only with less than 4K context.
No
Will not load on that machine at any published version.
Speed is generation speed for one user (batch 1) at the stated context.
Image and video models
Kingy times one standard output: a 5-second clip at the model’s native resolution, or one 1024×1024 image.
Verdict
One 5-second clip
One image
Yes
5 min or less
30 s or less
Yes, but slow
5–30 min
30 s to 3 min
Barely
Over 30 min, only at reduced resolution or length, or with heavy offload (ComfyUI partial offload counts)
Over 3 min, or with offload
No
Won’t load
Won’t load
Media speed depends too much on compute and pipeline settings to estimate, so media verdicts are always measured or reported. Without either, they read “Not tested yet”.
How Kingy knows
Kingy measured: Kingy ran it. Each result has a Local Lab receipt with the exact file and hash, runtime build, command, context, speed and date. Failed attempts are kept too.
Reported: a named third party ran it. Kingy links the report and records their machine, file, runtime and speed. “Runs fine” without numbers does not count.
Estimated: calculated, not run, from the published file size and the formula below. An estimate always says “(estimated)” and never shows as a bare “Yes”.
No evidence yet: no measurement, no reliable report and, for text models, no published weight size to estimate from. The verdict is then “Not tested yet”. Kingy never invents a verdict to avoid saying so.
It fits if the memory needed is no more than the machine type’s usable memory.
Speed ≈ 0.5 × memory bandwidth ÷ bytes read per token, where bytes per token = active parameters × bits per weight ÷ 8. Mixture-of-experts models use active, not total, parameters. Real runtimes usually reach 60–80% of bandwidth, so 0.5 errs on the slow side.
On GPU PCs, if more than 25% of the weights must sit in system RAM, the estimate is Barely at best.
An estimated “Yes” needs at least 40 tokens/s, twice the threshold, and 20% spare memory. Otherwise the best an estimate can say is “Yes, but slow”. An estimated “No” means even the smallest published version is bigger than the machine’s usable memory plus, on GPU PCs, the system RAM. When a measurement and an estimate disagree, the measurement wins.
Machine types
Verdicts are per machine type, not per exact computer. Each type has a reference machine, and estimates use its numbers. Newer chips are often faster, so estimates err on the slow side. GB means 10⁹ bytes.
Machine type
Reference machine
Memory a model can use
Memory bandwidth
16GB laptop
MacBook Air M4, 16GB
10.5 GB (macOS lets the GPU use about 2/3 of RAM)
120 GB/s
PC with a 24GB GPU
RTX 4090 24GB with 64GB DDR5 system RAM
22.5 GB VRAM (24 minus 1.5 for driver and display)
Every verdict names the file tested or estimated, its format, the runtime, and how it compares with the official release. Local does not automatically mean compressed, and online does not automatically mean full quality.
Measured and reported verdicts are marked “Recheck due” after 90 days, estimates after 30 days. The old result stays visible until the recheck.
Use the data
The same records feed the verdict box in each Kingy post, the can_i_run_locally tool on the Kingy Facts MCP server, and the public API: