AI News

Local AI Models Tested: 8 Models, 120 Graded Tasks, Scored by Code

By Kingy AI · Reviewed by Curtis Pyke. Figures from Kingy AI’s own tests, run November 7 to 12, 2026.

Which free AI models are worth running on your own computer? We gave 8 open models the same 120 tasks, three times each, on a 128 GB mini PC. A script marked every answer against a known right answer. No AI judged another AI.

The short answer

  • Kingy AI’s home model test (November 2026) found that 5 of 8 open models tied at the top of a 120-task test, scoring 93% to 98%, based on 360 marked answers per model.
  • Kingy AI’s home model test (November 2026) found that Llama 3.3 70B scored 76%, behind Gemma 4 26B at 93%, a model about a third of its size, based on 360 marked answers each.
  • Kingy AI’s home model test (November 2026) found that switching thinking on lifted Gemma 4 26B from 93% to 98% and raised chip energy from 143 to 2,249 watt-hours per 1,000 answers.
  • Kingy AI’s harder test set (November 2026) found that open models on home machines scored 38% to 65%, against 97% and 100% for two top cloud models, based on 120 tasks run once.

Results: the first test set

The first set has 120 tasks: 60 everyday jobs (dates, money, schedules, lookups), 30 coding tasks and 30 writing tasks with rules to follow. Each model did every task three times, in its maker’s default mode with its maker’s settings.

ModelModeScoreLikely rangeEverydayCodingRulesTypical everyday answerTokens a secondChip energy per 1,000 answers
gpt-oss-120breasoning medium98%94 to 10099%97%99%17 s33.7593 Wh
Qwen3.6 35Bthinking on98%94 to 9999%94%100%48 s61.42,908 Wh
Qwen3.8 27B (3-bit file)thinking on97%92 to 9998%92%99%49 s14.78,087 Wh
Gemma 4 31Bthinking off96%91 to 9896%94%99%23 s10.4789 Wh
Gemma 4 26Bthinking off93%87 to 9795%92%91%5 s55.8143 Wh
Qwen3 8Bthinking on80%72 to 8687%78%67%37 s37.82,912 Wh
Llama 3.3 70Bno thinking mode76%67 to 8376%66%87%39 s5.01,555 Wh
Qwen3 4Bthinking on69%60 to 7777%61%62%28 s58.61,841 Wh

Kingy measured on the Lucy AI Studio, November 7 to 12, 2026. “Likely range” is the 95% range around the score. When two ranges overlap, the test cannot tell the models apart, and we call it a tie. A token is a piece of a word; ten tokens is about seven words.

What the results say

  • The top is a tie. gpt-oss-120b, Qwen3.6 35B, Qwen3.8 27B (3-bit file), Gemma 4 31B and Gemma 4 26B have overlapping ranges. On this set, none is clearly better than the others.
  • Bigger is not better by itself. Llama 3.3 70B, released in December 2024, scored 76%. Every other model of 26B or larger scored 93% or more.
  • Small models drop off. Qwen3 8B scored 80% and Qwen3 4B scored 69%. Both are clearly behind the top group.
  • The biggest model was also quick. gpt-oss-120b gave a typical everyday answer in 17 seconds and used 593 Wh per 1,000 answers. Qwen3.8 27B thinks at length by default: 49 seconds and 8,087 Wh.

Thinking: better scores, much more time and power

Some models can think before they answer. They write hidden notes first, then the reply. We ran Gemma 4 26B both ways on the same machine. With thinking off, its maker’s default, it scored 93% and a typical everyday answer took 5 seconds. With thinking on it scored 98% and took 50 seconds. Chip energy rose from 143 to 2,249 Wh per 1,000 answers.

With thinking on, 4 of its 360 answers ran out of room before finishing. Those count as wrong.

The same model on a gaming laptop

We ran the identical Gemma 4 26B file on a gaming laptop with 8 GB of graphics memory. It scored 94% (likely range 88 to 97), a tie with the 93% on the mini PC. The answers are as good; they arrive more slowly. The laptop wrote 30.8 tokens a second against 55.8, because the model is bigger than the laptop’s graphics memory and part of it runs on the processor.

A harder set: home models against cloud models

The first set is too easy to rank the best models. So we built a harder set of 120 tasks: longer chains of everyday sums, coding across several files, writing with up to 12 rules at once, and questions with no right answer, where the honest reply is to say so. Each model ran it once.

ModelWhere it ranScoreLikely range
GPT-6.1 TerraCloud100%97 to 100
Claude Opus 5.5Cloud97%92 to 99
Qwen3.6 35BLucy AI Studio65%56 to 73
Gemma 4 31BLucy AI Studio64%55 to 72
Gemma 4 26BLucy AI Studio58%49 to 67
Gemma 4 26BGaming laptop55%46 to 64
Qwen3 4BGaming laptop38%29 to 46
  • The gap is clear. Every home model’s range sits below both cloud models’ ranges.
  • Read Qwen3.6 35B’s score with care. 35 of its 42 misses were answers cut off at its 32,768-token limit while it was still thinking. A higher limit might raise its score.
  • Claude Opus 5.5 left 3 tasks unanswered because the run reached its spending cap. Those count as wrong here.
  • This set is a trial version. It ran once per model, and we may change it before we use it for regular reports.

Top cloud models on the first set

For comparison, three cloud models did the first set through their paid APIs on November 9, 2026: GPT-6.1 Terra 99%, Gemini 4 Argon 99%, Claude Opus 5.5 99%. All three tie with the top home models. On tasks like these, a good home model and a top cloud model score the same.

How we test

  • Every task has a right answer that a script can check: a number, a date, code that must pass tests, or text that must follow stated rules.
  • No AI marks another AI’s work.
  • About 30% of the tasks are kept private, so future models cannot learn the answers. They are scored the same way.
  • Each model runs in its maker’s default mode with its maker’s recommended settings.
  • Each model file is fixed by its fingerprint (SHA-256), so the result belongs to that exact file. A more or less compressed copy of the same model can score differently.
  • An answer that is cut off, or missing, counts as wrong.

The machines

  • Lucy AI Studio, from Cephalon. AMD Ryzen AI Max+ 395, 128 GB of shared memory, Linux. Cephalon sent Kingy AI this machine for free and paid for an earlier video on our YouTube channel. Cephalon did not pay for this test and did not see the results before they were published.
  • Gaming laptop. NVIDIA RTX 5070 Laptop with 8 GB of graphics memory, 32 GB of memory, Windows 11. Kingy AI bought this machine.

Limits

  • Answers cut off by the output limit, out of 360: Gemma 4 26B with thinking on 4, Qwen3.6 35B 6, Qwen3 4B 2, Qwen3.8 27B (3-bit file) 6.
  • Qwen3.5 122B was on the list but its file failed to load, so it has no score.
  • One everyday task about splitting a bill can be read two ways. It is still in every score.
  • Energy is what the graphics chip reports, not power at the wall. A whole computer uses more.
  • Kingy AI’s research and writing are done with Claude, which also built these tests and started the runs. Claude Opus 5.5 is one of the cloud models compared. A script, not Claude, marked every answer.