Apple side-by-side human evaluation
Apple's in-house graders compare two models' responses to locale-specific text and image prompts and pick the preferred one. Results are win, tie and loss fractions per comparison.
Categories include analytical reasoning, chat, coding, math, summarization and tool use across several locale groups. The 2026 post compares against Apple's own 2025 models, while the 2025 post compares against open and closed models.
- Level
- Everyday Most adults, without training
- Scored
- Ranked by votes People pick the better of two answers, as in an arena, and the votes are turned into ratings.
- Answers
- Open-ended Free-form text, documents, designs or media judged as a whole.
AFM 3 Cloud leads with 64.7%
Best score per model as labs report them in their own cards. Labs pick their own settings, so these compare less cleanly than independent runs.
- 1 AFM 3 Cloud 64.7%
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | AFM 3 Cloud | 64.7% | Jun 2026 | Model card |
Scores over time
All the scores we have, one dot each, so a model tested at several settings appears more than once.
Labs headline it less often now, last in Jun 2026
Launch posts from every lab that has listed it, one tick per post. Tall colored ticks include it, and short gray ones leave it out.
| Date | Launch post | Score | As the post writes it |
|---|---|---|---|
| 8 Jun 2026 | AFM 3 Cloud | 64.7 | Side-by-side human evaluation, general text, AFM 3 Cloud preferred vs 2025 AFM Server (8.7 percent preferred the 2025 model) |
| 9 Jun 2025 | Apple Foundation Model Server (2025) | Fraction of preferred responses in side-by-side evaluation of text responses (chart only) |
One system card reports it
Besides launch posts, we read the system cards, model cards and technical reports labs publish with a release. They report far more, including safety and frontier-risk tests.
| Date | Document | Score | As the document writes it | Section |
|---|---|---|---|---|
| 17 Jul 2025 | Apple Foundation Model Server (2025) technical report | Side-by-side human preference, win/tie/loss per rival, no overall. Text EN: vs Llama-4-Scout 22.7/59.6/17.7, vs Qwen-3-235B 20.7/54.8/24.5, vs GPT-4o 16.7/68.3/15.0. Image EN: vs Llama-4-Scout 23/51/26, Qwen-2.5-VL-32B 31/44/26, GPT-4o 15/50/35. EN outside US and PFIGSCJK also given (Figures 4 and 5) | Capabilities |
More in Instructions and chat
- IFEval In use Models answer about 500 prompts containing verifiable instructions such as word limits or required keywords. Code checks each instruction for strict and loose accuracy.
Gemma 4 31B IT 98.9% - IFBench In use 58 new out-of-domain verifiable output constraints. Scored by programmatic constraint checks (strict or loose).
Nemotron-Cascade 2 30B-A3B 82.9% - MultiChallenge In use Multi-turn conversations testing instruction retention, inference memory, versioned editing and self-coherence. Graded by LLM judge with rubrics. GPT-6.1 Sol 82%
- Multi-IF In use Models follow verifiable instructions over three-turn conversations in eight languages, extending IFEval. Scored by instruction accuracy at each turn. LFM3-1.7B 89.1%
- AdvancedIF In use Models respond to over 1,600 prompts with complex, multi-turn, and system-level instructions. A judge model checks each response against expert-written rubrics.
MAI-Thinking-1 85% - IFStruct In use Models generate JSON or YAML that must follow a requested schema and phrasing constraints, without constrained decoding. A response passes only if every structural constraint holds. LFM3-1.7B 93.7%