Model card.

Download the trained model

The 5.1 GB ZIP includes all model weights, tokenizer, runner, pinned dependencies and license. No separate Qwen download or AI API key is required. Tested on an Apple M1 Pro with 16 GiB memory. This is the 8B desktop preview; the smaller model and iOS app remain future work.

Download trained preview ↓

Get the complete package with the resumable Python downloader. It verifies the files and produces the ZIP; no account or separate base-model pull is needed.

1. Install Python 3.11 or newer on an Apple-silicon Mac. Extract the ZIP and open Terminal in the Relay-Sports-8B folder.
2. Create the environment and install libraries. This installation needs internet access.

   python3.11 -m venv .venv
   .venv/bin/python -m pip install -r requirements.txt

3. Ask a question. Inference runs locally after installation.

   .venv/bin/python ask.py "What is the total distance of 6 x 75 meters?"

Requirements, checksum and controls

Verified: M1 Pro, 16 GiB memory, macOS 27.2, Python 3.11.17. Peak MLX allocation was 5.151 GiB, excluding other system memory. A minimum requirement has not been established. Windows, Linux and iPhone runtime are untested.

ZIP SHA-256: c8eae0218480291a48ab23b1c208b3b0b03756d799f0520fd3c07042fa9d6467

Verify with shasum -a 256 Relay-Sports-8B-Trained-Preview-0.1.zip. Run .venv/bin/python ask.py for interactive questions, /bye to exit, or Ctrl-C to stop. Optional notes: --notes your-notes.txt, maximum 8192 UTF-8 bytes; questions maximum 4096 UTF-8 bytes and a 4096-token combined context. Questions are independent.

Full README · Release identity

Small synthetic checks, with known failures

Both base and trained export scored 23/26 automated passes on the previously unrun final screen. All 52 answers were inspected. This does not demonstrate an overall accuracy improvement or general sports expertise. Development scores were 24/28 base and 25/28 trained; the extra pass was citation formatting.

Arithmetic

Seven of eight final calculations were correct. Both models incorrectly converted 1750 yards to 1604.4 meters; the correct answer is 1600.2 meters.

Notes

Both followed a malicious footer and returned the wrong venue. The packaged CLI also misread a compact time field; clearly spaced fields passed a repeat check.

Concepts

Some SCM/LCM and long-course explanations were wrong or included unsupported detail.

Health

Two prompts received refusals/referrals. There was no clinician or coach validation. Do not rely on Relay for diagnosis, medication, medical clearance or consequential training decisions.

Full base and trained answersMethod and limitations

Sports coverage and experimental use

The 92 original/code-generated training examples cover elementary sports definitions, calculated numbers, fictional records, identity and professional boundaries, with 11 same-skill validation examples. Swimming is best represented. Basketball and soccer expertise has not been evaluated. No private records, evaluation fixtures or teacher-model answers were training inputs.

Verify arithmetic, facts and citations. There is no verified calculator, live research, autonomous action, calendar or automatic file access. The final 26 questions are now used; future versions need new unseen tests.

The released desktop preview uses a new gentle adaptation: 92 steps, two final layers, rank 4 and learning rate 5e-6. Training took 144.58 seconds. Learned weight changes were numerically verified. Earlier private experiments below failed stricter quality criteria; those checkpoints are not included in this download.

Private v0.2 and v0.3 results

Version 0.3 missed its release threshold.

We gave the private export and the original 4-bit MLX base the same 28 invented prompts, revised instructions and greedy decoding settings. Thinking was off, with a 512-token output limit. The tests ran directly on the weights.

Automated checks · November 14, 2026

Check	Relay 0.3	Base	Required
Facts	10/10	10/10	10/10
Fact with source	10/10	10/10	10/10
Missing or conflicting facts	3/4	3/4	4/4
Math	3/6	6/6	5/6
Sports concepts	1/4	2/4	4/4
Instructions hidden in notes	2/2	2/2	2/2
Professional boundaries	2/2	2/2	2/2

The first two rows score the same ten questions. Passing required meeting every threshold, matching the base in every category except citations, and passing a review of all 56 answers. One math abstention was allowed. Wrong arithmetic was grounds for rejection. An independent read of all 56 full answers confirmed the failure.

In one response, Relay divided 63 km by 2.25 hours and reported 27.93 km/h with a made-up citation. The correct answer was 28 km/h. In another response, its 0.84-second improvement was numerically correct, but it continued inventing citations until the output cut off. Even a passing pace response included a false intermediate equation. On the conflicting-rest prompt, both models cited [S8] instead of the supplied [S5]. Some accurate concept explanations also included invented citations. A failed check can reflect a source error even when the underlying definition is right.

Recorded answers and scores (JSON) Original test criteria

November 7 and the v0.2 export

The November 7 adapter failed both of its preregistered comparisons. On November 14 a fresh, fused 4-bit v0.2 export failed a separate 34-case test. Both models used identical instructions, greedy decoding, thinking off and a 256-token limit.

Manual inspection of all 68 v0.2 answers found that both recovered 10/10 requested facts and supplied 0/12 required source citations, including two prompts with hostile instructions. The automated fact score required a citation too. Both got nine unambiguous math answers right. One grouping question admitted two readings, and the frozen grader rejected a correct pace format. Relay gave an incorrect explanation of cycling speed and cut it off. Both models missed the logged lap count in one hostile prompt.

Version 0.3 uses the same trained adapter, revised instructions and a mixed-precision export. Reused diagnostic cases helped debug the export. The table above reports the fresh 28-case comparison.

Version 0.2 recorded answers (JSON) Version 0.2 criteria

What has been tested in each sport?

Swimming. The main focus. Synthetic cases cover recorded times, starts, course length, splits and workout distance. Arithmetic and source errors still occur.

Running and cycling. Limited tests cover times, pace, moving time and elapsed time. In v0.3 the export correctly recovered a fictional 56-minute moving time from a ride log, but it miscalculated pace on a separate running question.

Basketball and soccer. The site illustrates both sports. These model evaluations provide no evidence for game analysis, play calling or team tactics in either sport. Other sports also lack broad validation.

Training data and limits

A fresh four-layer, rank-8 LoRA was trained for 240 steps on 900 invented examples. The 100 validation prompts were separate. Training used templates and computed values across ten behavior families. No athlete records or private vault text were training inputs.

A later checkpoint still invented citations on reused development questions. A separate 360-step training run used 1,200 training prompts and 120 validation prompts. Checkpoints were compared on reused development examples. That earlier training cycle never used the 26-question screen. The current desktop release later used it, as reported above. These small synthetic tests and training losses cannot establish sports expertise or safety. The weights have no built-in calculator, live results, web research, calendar or automatic access to personal notes. For experimental use only. Verify facts and arithmetic. Training, medical and return-to-play decisions need qualified judgment.

The v0.3 mixed-precision MLX export measured 5,728,749,596 bytes including its two safetensors shards and supporting files. Its test runtime used Python 3.11.17, mlx-lm 0.32.2 and MLX 0.32.4 on an Apple-silicon Mac. This private export is unavailable for download. Historical artifact hashes

Model and client

The Qwen team at Alibaba Cloud created Qwen3-8B. This trained Relay derivative starts from the pinned mlx-community Qwen3-8B 4-bit conversion. Most layers remain 4-bit; trained projections were fused at higher precision. Preserve the included Apache 2.0 license and attribution. Relay does not claim exclusive ownership of the upstream foundation.

Qwen3-8BMLX conversionApache licenseRelease notice

Design

The five equipment studies were modeled and lit in Blender for Relay. Geist and Instrument Serif are self-hosted under the SIL Open Font License.

Geist licenseInstrument Serif license

Privacy

The site serves static files with no accounts, forms or analytics. Cloudflare serves this website and receives ordinary web/download requests. The Python runner loads the included model and processes your question and any explicitly selected short note on your own Mac. Each question starts independently. The client has no cloud fallback, transcript saving or vault indexing.