Heretic-v2 quant of Qwen3 Swift 1.5 Flash ships with reverse-engineering-tuned imatrix for 96GB VRAM

jjusko20 · reddit · 2026-10-30

Two hobbyists released Swift 1.5 Flash Next Heretic-v2, an uncensored quant of Qwen 3.8 Swift 1.5 Flash specialized for reverse engineering and coding.

Key points:

jeffpeng3's v2 Heretic weights strip refusals while preserving reasoning and output length, a big step up from v1; the poster built the quant and imatrix recipes.
The imatrix corpus (837 entries, 400k tokens) weights chain-of-thought (22.2%), system-prompt adherence (19.5%), design/architecture (13.5%), agentic tool use (12.4%) and a new 8.8% reverse-engineering slice covering x86-64/ARM64 disassembly, Ghidra decompilation, ELF/PE/Mach-O internals, anti-debug, deobfuscation, CFG deflattening and Frida/eBPF tracing, plus 2% no-refusal content.
The new RE and direct-answer categories (43k tokens) were run through the Heretic weights themselves so the matrix captures actual model behavior.
Their matrix is fused per-channel with AtomicChat's general-purpose matrix: 70% custom where strong, sliding to the broad baseline where thin.
A custom dynamic quant targets exactly 96GB of VRAM with room for mmproj, 8192 prefill batching and KV streaming, 6GB smaller than Q4KXL.

Original post →

Heretic-v2 quant of Qwen3 Swift 1.5 Flash ships with reverse-engineering-tuned imatrix for 96GB VRAM

More from coding & agent