# Modification notices
DevelopingDad produced the original quantized export on October 8, 2026 from
Trial62 revision `568aa9764b4f695fb18881b97f6d4ee8b725f75c` and updated this
package on October 23, 2026.
- Original conversion: text MLPs packed as NVFP4; text attention weights
encoded as per-channel FP8; calibrated weight and FP8 KV scales added.
Tied BF16 target embedding/output head and media projections retained.
- October 23 checkpoint: 144 unused MLP activation-scale tensors removed;
all 1,245 retained tensors are byte-identical to the previous export.
Both config groups now disable activation quantization (W4A16/W8A16).
No additional training, weight merging or weight quantization occurred.
- Native tokenizer, processor, generation configuration and root chat
template remain unchanged. The optional response-contract template and
its builder are removed from the current package.
- A native-format compatibility template adds only system-block whitespace;
no instructions or custom prompt variables are included.
- The portable launcher now selects that compatibility template, 64K total context,
60 GiB FP8 KV, eight slots, prefix caching and long-prefill controls.
Thinking remains a caller-selected option. Automatic restart is disabled.
- Optional modified vLLM files retain upstream copyright/SPDX notices and
the isolated FP8 draft-head checks. Their code is unchanged by this update.
- Provenance, manifests, checksums, documentation and generic API diagnostics
updated to describe the weight-only checkpoint and current recipe.
