Myric/Flint-APEX-GGUF

Flint — APEX quants

Upload in progress. The 5 bpw GGUF is uploading now; 4 bpw and 3 bpw are queued behind it, followed by their recipes and the imatrix. This note will be removed once all three quants are present.

APEX quant ladder for RecursiveSI/flint, a 39.8B-total / ~2.5B-active MoE coding agent from Recursive (Apache 2.0). Weights are quantized with llama.cpp and evaluated with the 42-task OpenCode agentic suite.

Tiers

tier	file	size	effective bpw	status
5 bpw	Recursive-Flint-APEX-5bpw.gguf	23 GB	4.76	uploading
4 bpw	Recursive-Flint-APEX-4bpw.gguf	19 GB	3.95	queued
3 bpw	Recursive-Flint-APEX-3bpw.gguf	14 GB	2.99	queued

Recipes (Flint-5bpw.recipe, Flint-4bpw.recipe, Flint-3bpw.recipe) and the code-weighted imatrix are included as they finish uploading. The 4/3 bpw recipes use IQ expert types and the imatrix computed against the 5 bpw artefact; the 5 bpw tier was built structurally.

42-task result (5 bpw)

Full OpenCode agentic-coding sweep on the GB10, temp 1.0, top_p 0.95, top_k 20, context 262144, output cap 32768, reasoning high, one slot, compaction off.

suite	result
opencode_tasks	25/25
opencode_tasks_hard	5/5
opencode_tasks_frontier	9/9
opencode_tasks_tier4a	3/3
total	42/42 (100%)

Peak context used 120,829 / 262,144; zero voids and zero truncated slots. The result row and plots are in the APEX benchmark report.

Architecture

• 39.8B total / ~2.5B active decoder-only MoE
• 49 layers: 2 dense + 47 MoE; 256 routed experts (top-6) + 1 shared per MoE layer
• 35 sliding-window attention layers (window 513) + 14 full-attention layers
• GQA 32 query / 4 KV heads, head dim 128, per-head QK RMSNorm and learned attention sinks
• RoPE on sliding layers; no positional encoding on full-attention layers
• 262,144-token context, 150,144 vocab

Attribution

• Base model: Recursive, RecursiveSI/flint, Apache 2.0.
• Quantization and evaluation: this repository; llama.cpp.