Gnosis is a family of custom language models trained to operate in the RLM (Recursive Language Model) pattern: instead of attending over a long context directly, the model offloads that context into a variable inside a REPL and writes programs that slice, search, and recursively query it. The models run as the inference engine / orchestrator inside the prehend harness and are trained with the rlm-trainer stack.
One repository per specialty + size, with each release addressable by a git tag:
gnosis-lm/Gnosis-<Specialty>-<Size> # full-precision (bf16) weights, tag = version
gnosis-lm/Gnosis-<Specialty>-<Size>-W4A16 # compressed-tensors quant (vLLM/SGLang serving)
gnosis-lm/Gnosis-<Specialty>-<Size>-GGUF # GGUF quants (llama.cpp / LM Studio / Ollama)
Load a specific version with revision="<tag>". The main (bf16) repo holds the canonical fine-tune base; the quant repos hold the deployment builds.
Retrieval-grounded decomposition over offloaded corpora (needle-in-a-haystack, ICD-10), tuned to reduce confabulation without regressing real-knowledge accuracy or tool use. Current release: v0.4.
| Repo | Format | Use |
|---|---|---|
| Gnosis-MedPolicy-12B | bf16 safetensors | canonical fine-tune base; load with transformers, re-quantize from here |
| Gnosis-MedPolicy-12B-W4A16 | W4A16 compressed-tensors (g32, int4) | vLLM / SGLang serving (~8 GB) |
| Gnosis-MedPolicy-12B-GGUF | GGUF Q4_0 + Q8_0 | llama.cpp / LM Studio / Ollama |
Gnosis models are built to run inside the prehend RLM harness over offloaded long context. They are not general chat models and are evaluated on retrieval-grounded decomposition. Medical outputs are research artifacts, not clinical or billing advice.
Apache-2.0, inherited from the google/gemma-4 lineage.