AutoTrust

Frontier AI research lab. We build AI scientists & the platforms they run on — compressing the full research lifecycle from idea to impact.

[AutoTrust @AutoTrustAI]

Our 31B open model: 92.7% on HumanEval.

Close to GPT-4.5 and Kimi K2. Ahead of Claude Opus 4.
Zero HumanEval-specific tuning. 0.20% of parameters trained.
Multimodal vision fully preserved.

Small model. Clean training. Real results.

→ https://t.co/Uug0fNgSL1

[AutoTrust @AutoTrustAI]

Two new open models are live from AutoTrust AI Lab 🚀

Gemma-4-31B Fable-5 Distilled
+ GGUF release with multimodal projector

92.7% HumanEval pass@1
+15.9 points over official gemma-4-31B-it
Vision preserved
Built for local text + image inference

https://t.co/XkDLeelThl

[AutoTrust @AutoTrustAI]

🚀 Our first community model is live

GPT-OSS 120B Fable-5 Distilled — GGUF. The first community GGUF of a LoRA-tuned GPT-OSS 120B, converted from Apple MLX

Runs locally on llama.cpp, LM Studio & Ollama. 243 downloads in 24h 🙏

🔗 https://t.co/LmfW3y065a

#LocalLLM #LLM #Ollama

[AutoTrust @AutoTrustAI]

These releases are part of our broader open-model stack underneath PaperGuru and our upcoming ScienceGuru research agents.

Feedback, bug reports, evals, and community quantizations are all welcome:
https://t.co/XkDLeelThl

[AutoTrust @AutoTrustAI]

@huggingface Result:

coding ability goes up, vision benchmarks stay at parity with the base model.

Most coding fine-tunes damage multimodal performance.

This one was trained specifically to avoid that.

[AutoTrust @AutoTrustAI]

AutoTrust open model release wave #2 is live on @huggingface

We’re releasing two new models in the Fable-5 family:
autotrust/gemma4-31B-Fable-5-Distilled
autotrust/gemma4-31B-Fable-5-Distilled-GGUF

For agentic coding, multimodal reasoning, and scientific research workflows.

[AutoTrust @AutoTrustAI]

A distilled, coding + agentic-reasoning 120B MoE (128 experts, 4 active).

Getting MLX → GGUF working meant solving MXFP4 dequant, LoRA fusion & GPT-OSS metadata quirks in llama.cpp.

Now it just runs — Mac/Win/Linux, CPU/CUDA/Metal/Vulkan/ROCm.

#llamacpp #GGUF #opensource

[AutoTrust @AutoTrustAI]

🚀 Our first community model is live

GPT-OSS 120B Fable-5 Distilled — GGUF. The first community GGUF of a LoRA-tuned GPT-OSS 120B, converted from Apple MLX

Runs locally on llama.cpp, LM Studio & Ollama. 243 downloads in 24h 🙏

🔗 https://t.co/LmfW3y065a

#LocalLLM #LLM #Ollama

[AutoTrust @AutoTrustAI]

Frontier AI research lab. Building AI scientists + the platforms they run on.

Months of research → days.

First: @PaperGuru_AI — SOTA on PaperBench. 23 peer-reviewed papers. ICML · FSE · TOSEM.

https://t.co/9bp06TETGo | From idea to impact, autonomously.