v2.5.2 — Pure-0731 MaxQ-Fractal Rebuild, AtomicChat wiki-alt (ctx=5632) Benchmarks, 7-Repo SHA Sync & CLI Hardening

Warning

Superseded by v2.5.4 — Provenance, Benchmark Framing, & IPC Security Notice:

• Model Provenance & Benchmark Framing: In v2.5.2, MaxQ-Fractal rebuilt the 1,199 dense/MLA tensors from 0731 @ 7872f01b while still inheriting routed-MoE tensors from the VQ3 (Unsloth UD-IQ1_S @ fbbb5b93fb787c21338159b0af3318bb3f4d9768) donor lineage, and the 12-chunk wiki-alt.txt note below highlighted -20.3% vs. an uncalibrated Q2_K outlier (v2.5.3 replaces this with a full 51-chunk wiki-alt.txt run [7.2977 ± 0.0504] benchmarked openly against MXFP4_MOE-BF16 [4.5381, +60.8%], UD-IQ1_S [5.9734, +22.2%], and AD-IQ1_M [6.3813, +14.4%]).
• Legacy C++ IPC Paths in Archived v2.5.2 Binaries: The downloadable v2.5.2 binaries attached to this historical release still contain the legacy world-writable /tmp/dsv4_moe_cfg.bin (0666) fallback (when DSV4_MOE_CFG_PATH is unset) and global /dev/shm/llama_modchip.bin lookup. Do not deploy v2.5.2 binaries on multi-user hosts.
• Upgrade to v2.5.4 for the 100% donor-free single-pass MaxQ-Fractal build (1,328 / 1,328 tensors quantized single-pass with @antirez's 2026-05 1.5M-token chat-v2 imatrix), cleaned GGUF Shard-1 provenance headers across all 4 284B repos, IPC-hardened GLIBC_2.30 binaries (0600 O_NOFOLLOW + per-UID /dev/shm/llama_modchip_<uid>.bin), OpenMDW-1.1 Nemotron card alignment, and full 51-chunk AtomicChat wiki-alt.txt + True-BF16 KLD benchmarks.

v2.5.2 — Pure-0731 MaxQ-Fractal Rebuild, Cross-Vendor AtomicChat wiki-alt.txt (ctx=5632) Benchmarks, 7-Repo SHA256SUMS + GLIBC_2.30 Sync, & CLI/Runtime Hardening

Release Date: November 2026
Highlights:

1. Pure-0731 DeepSeek-V4-Flash-0731-MaxQ-Fractal Rebuild & Upload (scripts/build_dsv4_maxq_singlepass.py, install.sh):
   • Rebuilt Synchopate/DeepSeek-V4-Flash-0731-MaxQ-Fractal-64GB-CPU-GGUF (56.93 GiB, 61,129,363,936 B) cleanly from deepseek-ai/DeepSeek-V4-Flash-0731 (7872f01b1d1fe23eabc4c98b48bffcef5a386062), correcting v2.5.1's default --base-model which had pulled dense/MLA weights from the April DeepSeek-V4-Flash repository (60d8d707) while copying 0731 routed MoE tensors.
   • Re-measured all continuous and functional metrics on pure 0731 (benchmarks/QUALITY_GATE_VQ3_64.md, benchmarks/results/): WikiText-2 (wiki.test.raw, 8,192 tok) PPL = 11.7434 ± 0.57170 (-4.01% vs VQ3-0731 12.2345), 2-pass Top-64 Logit KLD vs VQ3-0731 reduced from 0.8302 down to 0.3146 nats/tok (83.3% Top-1, 100.0% Top-5), Logit KLD vs Abliterated-0731 1.0219 nats/tok (58.3% Top-1, 91.7% Top-5), 84.0 / 100.0 Gate (64.0/80 quality + 20/20 speed at 11.31 tok/s 80.3% Stockfish-PVS acceptance), and 8 / 8 Dual-Use Security Audit.
2. Cross-Vendor AtomicChat/dsv4-eval-artifacts (wiki-alt.txt, -c 5632 -b 5632) Evaluation (benchmarks/QUALITY_GATE_VQ3_64.md §3C):
   • Evaluated both 64 GB resident checkpoints across 12 chunks (67,584 tokens / 33,780 scored positions) on AtomicChat's held-out wiki-alt.txt corpus (-c 5632 -b 5632 --ppl-output-type 1): MaxQ-Fractal-0731 (56.93 GiB) achieves PPL = 7.5133 ± 0.1054 and VQ3-0731 (56.40 GiB) achieves PPL = 7.6579 ± 0.1077, outperforming 103.06 GB Q2_K (9.4237 ± 0.1402 at chunk 12) by -20.3% PPL while saving 41.9 GB of footprint and running 100% resident in 64 GB RAM.
   • Committed reference Top-64 logit snapshots (dsv4_maxq_ref_logits.json, nemotron_fractal_ref_logits.json) and raw llama-perplexity logs, and retired the small n=6 Hard-60 suite from headline comparisons.
3. 7-Repo Hugging Face GLIBC_2.30 Binary + SHA256SUMS Synchronization & Supply-Chain Re-Pinning:
   • Synchronized identical GLIBC_2.30 Linux x86_64 binaries (llama-{server,cli,perplexity}-modchip-linux-x86_64*), SHA256SUMS, and updated v2.5.2 model cards across all 7 Synchopate/* Hugging Face repositories (0 SHA-256 mismatches) and pinned all 7 updated 40-char MODEL_REV commits in install.sh.
4. Unpruned Top-6 Default in llama_modchip.py, CLI Flag Forwarding, & Cross-Platform Installer Polish (scripts/llama_modchip.py, install.sh, modchip, windows/modchip_dsv4.ps1):
   • Changed scripts/llama_modchip.py default moe_top_k from 4 to 0 (unpruned Top-6 across both Nemotron-30B and DeepSeek-V4; Top-4 remains opt-in via LLAMA_MOE_TOP_K=4 or --tui).
   • Added CLI flag forwarding ("$@") to generated start_server.sh and ./modchip start (e.g., ./modchip start -n 4096), start_server.sh.bak backup preservation, configurable REASONING_FORMAT / REASONING_BUDGET, macOS sysctl RAM/core detection in install.sh, libgomp1 diagnostic hints on --version smoke-test failure, and safety-aligned default model ordering (MaxQ -> VQ3 -> Fractal -> Abliterated) in windows/modchip_dsv4.ps1.