Archived construction records. Historical candidates mentioned below are not all released. release-manifest.json identifies this release exactly.

Swift 1.5 IQ2_XS at ISTA parity (2026-09-24 ~01:42-01:45 CEST). KLD vs Swift-1.5 BF16, wiki.test.raw -c 512 --chunks 100.
ISTA release IQ2_XS (vs base Qwen BF16): 0.193998
 V1MIX    : 0.191398 +-0.002515  file Swift15-GSQ-V1MIX-ISTAalloc-IQ2_XS.gguf
 ISTAIMAT : 0.193108 +-0.002524  file Swift15-GSQ-ISTAIMAT-ISTAalloc-IQ2_XS.gguf
Recipe (both): ISTA-DASLab Qwen3.8-27B-GSQ-RCO IQ2_XS per-tensor allocation forced via llama-quantize --tensor-type-file (llama.cpp fee39dd)
 imatrix: V1MIX = v1 make_calib.py corpus (fineweb-edu 30/OpenThoughts3 30/multilingual wiki 20/code 12/Platypus 8, raw text), 309 chunks x 4096, 8 shards merged;
          ISTAIMAT = ISTA published imatrix-qwen3.8-27b.gguf
 refinement: refine_gguf_fast.py (checkpoint-commit fix, ORIGINAL objective: IQ1 neighbours dense), tokens_2048.pt (same v1 corpus, 512x2048),
             --n-train 512 --n-val 32 --batch-size 8 --micro 4 --epochs 3 --no-vq-codes --lr-scale-iq 1e-3, TileLang backward (nvcc on PATH)

IQ2_S (ISTA release 0.135091): V1MIX 0.134751 +-0.001950 ; ISTAIMAT 0.135813 +-0.002045 — same recipe as IQ2_XS above.
IQ3_XXS (ISTA 0.098121): V1MIX 0.097817 +-0.001484 ; ISTAIMAT 0.098618 +-0.001556 — same recipe.
-mtp files: add_mtp15.py (refined tensors byte-identical + blk.64 Q6_K/F32), MTP_OK verified.
IQ3_S (ISTA 0.051851): V1MIX v1 recipe 0.051265 +-0.000764 (-1.13%); ISTAIMAT 0.052859.


IQ2_XS FIXED-OBJECTIVE (Astra IQ1 stage_objective patch) v1 recipe, v1mix imatrix + ISTA allocation: KLD 0.189979 +-0.002502 (ISTA 0.193998, -2.07%)
IQ3_XXS FIXED-OBJECTIVE v1 recipe: 0.097774 +-0.001484 (ISTA 0.098121, -0.35%)