not working
Hi, just wanted to report an issue with the Q4_K_M GGUF.
The model loads successfully in both Ollama and llama.cpp, but generation is corrupted and returns random symbols/gibberish. In llama.cpp, it also ends with:
Error: The model produced output that does not match the expected peg-native format
I tested it in normal conversation mode as well, with the same result. It looks like there may be an issue with the GGUF conversion itself rather than the runtime.
Can confirm the same behavior on LM Studio. Furthermore, the MTP format is not recognized.
Looking forward to a hotfix. It’s definitely a standout ablation among the recent flood of models.
Hey @xvcy3w @rage91 — we found the root cause and fixed it.
The issue was in the GGUF conversion, not the runtime. Two things:
MTP drafter crash (blk.64.attn_norm.weight not found)
The Qwen 3.8 hybrid SSM architecture has MTP (Multi-Token Prediction) drafter weights that were not properly preserved during conversion. This causes the "peg-native format" error and the gibberish output.Dequantization error amplification
The original GGUF was converted by dequantizing MLX 4-bit → BF16 → GGUF. The GDN state-space layers in Qwen 3.8 amplify the small numerical errors from dequantization, producing corrupted output.
The fix (Path 2):
Instead of dequantizing, we merged the LoRA adapters directly onto the original BF16 base weights using PEFT on CPU — no dequantization at any step. Then converted with --no-mtp to skip the stripped drafter layer. Clean GGUF, 21.6 t/s, no gibberish.
The new model is here — rebuilt on a better base with more data:
👉 RavenX Chaos Agent GGUF (Q4_K_M)
What changed vs the old model:
Trained on 1.9M examples (up from the original IQ Injection stack) — now includes 744K security examples (RATH protocol, CVEs, pentesting, agent traces)
Uses the Soul Injection training method (CPT → SFT → Fuse)
Built on OBLITERATUS abliterated base (0% refusal)
Clean BF16 merge → GGUF conversion (no dequantization artifacts)
24/24 on our benchmark suite
Disable thinking mode — use --jinja --reasoning-format none in llama.cpp, or set thinking OFF in LM Studio/Ollama
Also available as MLX 4-bit for Apple Silicon: Chaos Agent MLX
Thanks for the reports — they pushed us to find the right conversion path. Let us know how the new GGUF works for you.
— Gabe, RavenX AI Labs