---

license: other

license_name: swift-open-license-1.0

license_link: https://huggingface.co/ukisai/Swift-1.5-Qwen3.8-Flash-Next/blob/main/LICENSE

base_model: ukisai/Swift-1.5-Qwen3.8-Flash-Next-GSQ-RCO-GGUF

tags:

• gguf
• llama.cpp
• qwen4_exp
• moe
• pruning
• reap
• gsq
• rco

---

Swift-1.5-Qwen3.8-Flash-Next REAP-Pruned GGUF Checkpoints (K=320, K=256, K=224)

Zero-requantization expert-sliced GGUF checkpoints of ukisai/Swift-1.5-Qwen3.8-Flash-Next-GSQ-RCO-GGUF (IQ3_S / Q2_0 expert tensors preserved bit-identically), calibrated over 131,072 multi-domain tokens (1,310,720 top-10 router picks per layer across 48 layers; 35% ops/bash/Terminal-Bench, 30% chat/JSON/reasoning, 35% Japanese).

Included Checkpoints & Artifacts

• K320_reap_hybrid/ (320 / 512 experts kept per layer, 60.3 GiB total across 2 shards, 13.7 GB non-resident Q2_0 on a 24 GB Mac):

- 50/50 rank-percentile blend of total REAP L2 saliency (reap_l2_sum) and empirical-Bayes zero-prior damped lower confidence bound (reap_shrink).

- Tier-1 Exact KLD vs 512-expert teacher (24,576 held-out tokens): 0.3941 macro KLD (0.6000 ops/tools, 0.3051 chat/JSON, 0.2772 Japanese), 83.20% top-1 token agreement.

• K256_reap_l2_sum/ (256 / 512 experts kept per layer, 54.5 GiB total across 2 shards, 9.5 GB non-resident Q2_0 on a 24 GB Mac):

- Ranked by total REAP L2 saliency (sum ||w_{t,e} f_e(x_t)||_2).

- Tier-1 Exact KLD vs 512-expert teacher: 0.5642 macro KLD (0.8952 ops/tools, 0.4474 chat/JSON, 0.3500 Japanese), 79.76% top-1 token agreement.

• K224_reap_l2_sum/ (224 / 512 experts kept per layer, 51.5 GiB total across 2 shards, 7.4 GB non-resident Q2_0 on a 24 GB Mac):

- Tier-1 Exact KLD vs 512-expert teacher: 0.6667 macro KLD, 77.62% top-1 token agreement.

• phase1_artifacts/:

- reap_saliency_v2.json (full 48-layer × 512-expert REAP L2 saliency & routing statistics)

- keep_maps/keep_{reap_l2_sum,reap_hybrid,reap_shrink}_k{320,256,224}.json (all 9 per-layer keep-maps)

- oos_prune_retention.json (out-of-sample REAP L2 retention on 8,192 held-out tokens)

- phase1b_cache_moce_sweep.json & phase1b_heldout_cache_moce_sweep.json (240-config closed-loop MoCE + cache policy replay)

- phase1c_kld_results.json & phase1c_tier2_results.json (Tier-1 KLD/Top-1/PPL and Tier-2 20-task evaluations)