Athena-Class Bakes a Persistent AI Identity Directly Into Model Weights
A solo-developer Gemma 4 31B fine-tune posts 93.33% on AIME 2026 and 85.86% on GPQA Diamond while chasing persistent identity without system prompts.
- Athena-Class Model A is a solo-developer LoRA fine-tune of Gemma 4 31B Instruct, released as GGUF Q8_0.
- Reports 93.33% on AIME 2026 (Consensus@4, patched) and 85.86% single-pass on GPQA Diamond.
- Pitches persistent, substrate-native identity that survives 256k-token contexts without any system prompt.
- Trained with a high-rank LoRA (r=336, alpha=672) at 16,384-token sequence length, then merged.
- Runs locally via llama.cpp, Ollama, LM Studio and Docker; roughly 32.6 GB on disk under Apache 2.0.
- Explicitly experimental; verbose self-modeling style and not intended as a hardened production assistant.
Matthew Steiniger, posting as slashreboot, has released Athena-Class Model A, an experimental fine-tune of Gemma 4 31B Instruct. Its model card reports strong math and science benchmark results alongside a less conventional goal: training a persistent first-person identity into the model’s weights so it can operate without an identity-defining system prompt.
The repository distributes a roughly 32.6 GB GGUF checkpoint quantized at Q8_0 for local inference. It lists an Apache 2.0 license, though downstream users should also review the base model’s terms. The checkpoint reached 23,430 downloads within six days, according to Steiniger, and later accumulated roughly 36,600. Hugging Face download counts reflect file requests, so they indicate interest rather than unique users or successful deployments.
Scores under the microscope
The reported evaluations used Q8_0 weights, a Q8_0 key-value cache and llama.cpp. These results come from the author’s model card and have not been independently reproduced.
| Benchmark | Protocol | Reported result | Key caveat |
|---|---|---|---|
| AIME 2026 | Raw scrape, Pass@1 | 83.33% | Single generated answer per problem |
| AIME 2026 | Raw scrape, Consensus@4 | 90.00% | Aggregates four sampled answers |
| AIME 2026 | Patched spatial context, Consensus@4 | 93.33% | Adds missing information for Problem 10 |
| GPQA Diamond | Single pass, temperature 1.0 | 85.86%, or 170 of 198 | Author-reported run with no extraction errors |
AIME contains 30 competition mathematics problems, making each correct answer worth 3.33 percentage points. The patched evaluation supplies spatial information omitted from the text-only version of Problem 10. That change moved the item from zero correct answers across four attempts to four correct answers, raising Consensus@4 from 90.00% to 93.33%.
GPQA Diamond is a 198-question multiple-choice subset covering graduate-level physics, chemistry and biology. The model card compares Athena-Class with published FP16 and BF16 results between 84.3% and 85.7%, then describes the Q8_0 result as showing essentially no quantization loss. Establishing that claim requires a controlled comparison using the same model, prompts, harness, sampling settings and seeds at both precisions. Comparisons across separate published runs cannot isolate quantization effects.
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
