GestaltLabs/Ornstein3.6-35B-A3B-RYS-SABER-GGUF

chat_template: robust to any client (Claude Code, OpenCode, Pi, etc.). Merges multi/mid-conversation system messages, drops no-user-query hard-fail, tolerates unknown content shapes, unknown roles, missing tool args, system+vision combos.

README: --reasoning-format deepseek is default in recent llama.cpp; --jinja + --chat-template-file is enough

README: note that thinking model needs bundled chat_template.jinja + --jinja --reasoning-format deepseek to hide <think> tags

Add Qwen3-Thinking chat_template.jinja to fix raw <think> tags in llama-server output

Upload Q3_K_M (RYS-patched)

Upload Q4_K_M (RYS-patched)

Update README: RYS layer-type metadata fix, patched llama.cpp fork, usage instructions

Upload Q5_K_M (RYS-patched)

Upload Q6_K (RYS-patched)

Upload Q8_0 (RYS-patched)

Clear out stale GGUFs (pre-RYS-patch convert_hf_to_gguf)

Add linked citations for prior art

Add Q6_K, Q5_K_M/S, Q4_K_M/S, Q3_K_M/S

Upload Ornstein3.6-35B-A3B-RYS-SABER-Q8_0.gguf with huggingface_hub

Upload Ornstein3.6-35B-A3B-RYS-SABER-F16.gguf with huggingface_hub

Upload ornstein3.6RYS-SABER.jpeg with huggingface_hub

Upload README.md with huggingface_hub

initial commit