OpenELM 3B
by Apple · United StatesApple's open research model, using layer-wise scaling to spend parameters where they matter most.
Specification
The numbers
- Maker
- Apple
- Released
- 2024-04
- Parameters
- 3B
- Architecture
- Dense transformer
- Context window
- 2,048 tokens
- Max output
- 2,048 tokens
- Input
- Text
- Output
- Text
- Reasoning
- No
- Tool calling
- No
- Knowledge cutoff
- Not reported
- Licence
- Apple Sample Code License
- Availability
- Not reported
- Weights
- Downloadable
Cost
Price per million tokens
- Input
- $0.01 / M tokens
- Output
- $0.02 / M tokens
- Blended 3:1
- $0.013
This model has open weights, so there is no first-party price. The figures above are a representative third-party hosting rate — you can also run it yourself for the cost of the hardware.
Published scores
Benchmarks
Figures published by Apple or taken from a public leaderboard. Row last checked 2026-08.
| Category | Score | Rank |
|---|---|---|
| ReasoningMulti-step logic on problems that cannot be looked up | Not reported | — |
| MathsCompetition mathematics, graded on the final answer | Not reported | — |
| CodingWriting and repairing real code | Not reported | — |
| KnowledgeMMLU-Pro | 20.0% | 77 of 77 reporting the same tests |
| MultimodalReading charts, diagrams and photographs | Not reported | — |
| Instruction followingObeying an exact, checkable format | Not reported | — |
| Human preferenceWhich answer people pick, blind | Not reported | — |
A category averages every benchmark in it that OpenELM 3B reports. The rank counts only models that report the same tests, so it never compares an average over three benchmarks against an average over one.
Every reported test
| MMLU-ProKnowledge | Rank 77 of 77 models reporting | |
|---|---|---|
| GPQA DiamondReasoning | Not reported | No figure published |
| AIME 2025Maths | Not reported | No figure published |
| MATH-500Maths | Not reported | No figure published |
| SWE-bench VerifiedCoding | Not reported | No figure published |
| SWE-bench ProCoding | Not reported | No figure published |
| Terminal-Bench 2.1Coding | Not reported | No figure published |
| Frontier-Bench v0.1Reasoning | Not reported | No figure published |
| Terminal-Bench 4.0Coding | Not reported | No figure published |
| LiveCodeBenchCoding | Not reported | No figure published |
| HumanEvalCoding | Not reported | No figure published |
| MMMUMultimodal | Not reported | No figure published |
| IFEvalInstruction following | Not reported | No figure published |
| LMArena EloHuman preference | Not reported | No figure published |
Where these numbers come from
Every score on this page is a published figure, taken from the model's own card, system card, technical report or release post, or from a public leaderboard. CorX Labs did not run these evaluations. Most are self-reported by the lab that built the model, which means they were produced under that lab's own choice of prompt, scaffold and number of attempts — so treat them as a starting point for a shortlist, not as a settled ranking.
A score someone other than the model's maker measured is marked Independent and names its measurer. Those are the stronger numbers on this page — an outside harness has no reason to flatter anyone — and there are not many of them.
Where a figure has not been published, the cell reads Not reported rather than an estimate. Nothing here is inferred, interpolated or guessed. Each model records the month its row was last checked. Full method and caveats.