Ornith-1.5 DFlash

Releases·Model Releasefeatured

Ornith AI releases Ornith 1.5 family: 9B, 35B-A3B, and 397B models

Ornith AI released the Ornith 1.5 family of open-source LLMs, spanning 9B dense, 35B MoE, and 397B MoE, trained with self-improving strategies. The models achieve state-of-the-art performance among open-source models of comparable size, with the 397B matching Claude Opus 4.8 on reasoning, agentic, and coding tasks. Community quantizations are already available for the smaller models.

Ornith AI released the Ornith 1.5 family of open-source LLMs, spanning 9B dense, 35B MoE, and 397B MoE, trained with self-improving strategies. The models achieve state-of-the-art performance among open-source models of comparable size, with the 397B matching Claude Opus 4.8 on reasoning, agentic, and coding tasks. Community quantizations are already available for the smaller models.

LiLiCorr is a lightweight likelihood-based model that correlates per-position marginal distributions from parallel drafters like DFlash, improving joint coherence of drafted token blocks in speculative decoding. It keeps top-k tokens at each position, enhancing inference speed without retraining the target model.

Social·Community

Users report Qwen 3.8 27B QAT Q2 runs in 13-14GB RAM with Q5 KV cache

A Reddit user highlights that Qwen 3.8 27B, quantized to QAT Q2, runs locally with only 13-14GB RAM usage when combined with a Q2 DFlash model and Q5 KV cache. The setup reportedly handles up to 200K tokens of context with minimal quality degradation, fitting on a 12GB GPU at reduced context.

A Reddit user highlights that Qwen 3.8 27B, quantized to QAT Q2, runs locally with only 13-14GB RAM usage when combined with a Q2 DFlash model and Q5 KV cache. The setup reportedly handles up to 200K tokens of context with minimal quality degradation, fitting on a 12GB GPU at reduced context.

Releases·Model Release

AMD releases Instella-MoE-16B-A3B open-source model on HuggingFace

AMD uploaded the Instella-MoE-16B-A3B model to HuggingFace on July 23, 2026, marking its entry into the open-source AI model space. The model uses a Mixture-of-Experts architecture with 16B total parameters and 3B active parameters per token.

AMD uploaded the Instella-MoE-16B-A3B model to HuggingFace on July 23, 2026, marking its entry into the open-source AI model space. The model uses a Mixture-of-Experts architecture with 16B total parameters and 3B active parameters per token.

Releases·Model Release

Nota AI claims REAPed 250B model beats DeepSeek-V4-Flash, prunes to 32B

Nota AI released Solar-Open2-250B-Nota-INT4-GlobalPruned on Hugging Face, claiming their 250B model outperforms max-think DeepSeek-V4-Flash. The model was REAPed down to a 32B variant, showcasing aggressive pruning while maintaining performance.

Nota AI released Solar-Open2-250B-Nota-INT4-GlobalPruned on Hugging Face, claiming their 250B model outperforms max-think DeepSeek-V4-Flash. The model was REAPed down to a 32B variant, showcasing aggressive pruning while maintaining performance.