Falcon-H1R Heretic V2 drops uncensored reasoning weights

This is a notable community model hack for hybrid architectures, but the core performance and safety claims are still mostly trust-based rather than independently validated.

• –The interesting technical bit is dual-layer targeting (`attn.o_proj` + `ssm.out_proj`) on a hybrid Falcon backbone, where many older abliteration workflows were built for pure Transformers.
• –Reported metrics (3% refusal, 0.0001 KL) come from a small self-reported test setup, so developers should treat them as early signals, not robust benchmark evidence.
• –The quantized footprint (~4.5 GB at Q4_K_M, per the post) makes this accessible for local experimentation on consumer hardware.
• –The release materially increases misuse risk by design; even with defensive intent, deployment needs strict sandboxing, logging, and access controls.