aether-models/aether-kit-olmo-2-1b-instruct

AetherKit OLMo 2 1B Instruct

A Core AI language bundle for AetherKit, ported as a v0.0.6 spike. Source: allenai/OLMo-2-0425-1B-Instruct at revision 48d788eca847d4d7548f375ad03d3c9312f6139e, Apache-2.0.

Folder	Platform	Weights / cache	Attention arithmetic	Context
mac/	macOS 27+ on Apple silicon	float16	float32	4,096 tokens

mac/ is a complete registered AetherKit bundle: language-model.json, model.aimodel/, tokenizer/, and export-report.json. The model is specialized by Core AI on first load. This repository contains the Kit registration format; use AetherKit's language products to open it.

Download and use

Download mac/** with Hugging Face tooling. Keep the complete directory structure, then pass the local mac directory to AetherLanguageModel:

from huggingface_hub import snapshot_download
snapshot_download("aether-models/aether-kit-olmo-2-1b-instruct", allow_patterns=["mac/**", "LICENSE", "NOTICE"], local_dir="olmo2-kit")

import AetherLanguage
import Foundation

let model = AetherLanguageModel(bundle: URL(filePath: "/absolute/olmo2-kit/mac"))
let session = AetherLanguageSession(model: model)
let response = try await session.respond(to: "Explain why the sky is blue.")
print(response.content)

For reproducible deployments, pin a repository commit as the download revision. AetherKit's catalog tooling can generate an entry with per-file sizes and SHA-256 values for ModelPreparation; this repo does not add automatic downloads to the runtime.

Qualification and limits

Exported with Torch 2.11.0, Core AI Torch 0.4.2, and Transformers 4.57.3. AetherKit source commit: 867b6fe0d56ed16079cfa987cef75297dcc0acde (the locally integrated v0.0.6 spike; remote main may not include the export tooling yet). Evidence in verification/ refers to exactly the model bytes recorded by the registration (model SHA-256 6e3aca1b961acf4f206db3684bd13e77f6f7043de2ab02b805b821d13f8d70ab).

• 32/32 native GPU argmax rows match an independent upstream float32 reference; minimum cosine 0.999999843 and all 64 greedy continuation tokens match.
• 1,024-token decoding completes with zero model reloads and bounded step shapes.
• Full-model CPU self-check, tokenizer chat cases, registered chat, streaming, seeded sampling, stop sequences, cancellation recovery, and FoundationModels adapter checks pass.
• Qualified on one M5 Max. iOS and other devices are unqualified. The extended native session gate exercised 2,048 tokens; the full registered 4,096-token context has not been qualified. This is an experimental spike bundle.
• Float32 attention avoids a reproduced fp16 CPU attention near-tie reversal; weights and cache storage stay fp16. No paired precision-performance benchmark is claimed.

License and conversion

See LICENSE and NOTICE. The checkpoint declares Apache-2.0 and provides no license file; the included license is the canonical Apache-2.0 text. The tokenizer files are unchanged. Weights were converted to float16 Core AI format. The upstream model card describes the model's intended uses and limitations.

The export report's local checkpoint path is redacted for publication; source repository, revision, hashes, versions, and conversion recipe are retained.