Repository navigation
v3.0 — Unified 3-Tier RAM Lineup (16 GB Nemotron-3.5 · 32 GB & 64 GB Qwen3.8-Flash-Next + DeepSeek-V4 284B), Zero-Collision ABI (types 44/45), 512-Bit Q1_0/Q2_0/GDN Engine, Zero-Copy 3D MoE --repack, & Static-libgomp GLIBC_2.30 Binaries
Release Date: November 2026
Highlights:
- Donor-Free
Qwen3.8-Flash-Next(MaxQ-64GB&MaxQ-32GBAligned) + Opt-InHeretic-2AbliteratedTwins & Complete 3-Tier RAM Matrix:MaxQ-64GB(57,969,748,352 B=53.9885 GiB,1,086standardggmltensors,max_type = 42,0custom types):- Shard 1 (
46,449,692,736 B=43.2596 GiB,SHA256 = f2b8bc79acb4b4d79033acd91c9c360ba96da92cd5841512befc450e15f4786f) + Shared Shard 2 (11,520,055,616 B=10.7289 GiB,SHA256 = 86b79714867743ddf5911ee7dff2717ae275f870551d86469cbdbf026c6b552e,per_layer_token_embd.weight [256, 320001536]inQ1_0 = 41) + optionalmmproj-Qwen3.8-Flash-Next-BF16.gguf(907,542,880 B,SHA256 = ef2af1d6aa8a22740c7252bc7ce40c5c2f1249aa8d9c8499a4d4572c3cfdf9df). - Compresses the
320M-row Per-Layer Embedding (PLE) table from26.82 GiB(IQ4_NL) to10.73 GiB(PCA-Wiener D=256 + RaBitQ Q1_0), reinvesting+16.09 GiBinto all512/512unpruned routed experts (K = 640),FP64 Woodbury DualGramsigned-scaleQ2_0, and the built-inblk.48NextN Multi-Token Prediction (MTP) module (0.81 GiB). - Blind
wiki-alt.txt(40 × 512chunks,0.000000%8/13-gram overlap with calibration) vs. Donor-FreeBF16Reference (Mean PPL(base) = 2.776624 ± 0.054868): AchievesMean PPL = 3.432500 ± 0.070829(3.43825 ± 0.07119underv3.0default--repack),Mean KLD = 0.497117 ± 0.009892,Median KLD = 0.105952, andSame top p = 78.314 ± 0.408%, outperformingISTA-DASLab GSQ-RCO-Coder-IQ1_M(54.397 GiB/58,408,584,928 B,256/512pruned experts, no MTP:Mean PPL = 5.091640 ± 0.121418,Mean KLD = 0.930241,Median KLD = 0.296442,Same top p = 69.382%) by-1.6591PPL (3.53×smaller PPL degradation overBF16),-46.6%Mean KLD, and-64.3%Median KLD while0.41 GiBsmaller.
- Shard 1 (
MaxQ-32GB(38,341,683,424 B=35.7085 GiB,1,128standardggmltensors,max_type = 42,0custom types):- Shard 1 (
26,821,627,808 B=24.9796 GiB,SHA256 = 06cb65f9cf6e510cd16500ca1719701a0ffa13dcaa4c4d0cc135e4f2071a8b4d,42ConeCollapsedK = 512layers with42 × F32ffn_down_exps.bias+7K = 640boundary layers +blk.48MTP) + Shared10.7289 GiBShard 2 (86b79714...) demand-paged via NVMemmap. - 32 GB Consumer Exclusivity (
18-Checkpoint Hugging Face Audit): Across all18publishedUnslothandISTA-DASLabQwen3.8-Flash-NextGGUFs (54.40–329.72 GiB),0 / 18fit inside a 32 GB host (<= 28.5 GiB).MaxQ-32GBpasses strict Linux Cgroup v2 enforcement (memory.max = 28.5 GiB,MemorySwapMax = 0,oom_kill = 0) atVmHWM = 26.91–27.44 GiB <= 28.5 GiBwith blindwiki-alt.txtMean PPL = 5.812001 ± 0.140197(5.80406repack) andMean KLD = 1.006916 ± 0.014944.
- Shard 1 (
- Opt-In Gated-Residual Over-Projected Abliterated Twins (
qwen38-64gb-ablit&qwen38-32gb-ablit,v3.0-r2):- Applies TPE-optimized signed multi-layer over-projection (
beta = 1.83 - 2.39acrossblk.31..47fromtrohrbaugh/Qwen3.8-Flash-Next-heretic-2, achieving0 / 100refusals at0.0818KLD by compensating forQwen3.8's widened multi-stream Gated Residual branches, plus signedbeta = 2.2508blk.48NextNMTP alignment so speculative drafting stays synchronized on uncensored completions) directly onto virginMaxQ-64GBandMaxQ-32GB(blk.0..30&token_embd100%untouched):MaxQ-64GB-Abliterated: Shard 1 =46,449,692,736 B(1,085tensors:27edited [16blk.32..47ssm_out/attn_output+5blk.31..35ffn_down_shexp+5blk.31..35512-expertffn_down_exps+1blk.48MTPattn_output],1,058bit-identical toMaxQ-64GB;SHA256 = 1dac27e81a3a73dcec21edea78f224eb48393a726e3f9ee7676c603d3174e82f) + Shared Shard 2 (11,520,055,616 B,SHA256 = 86b79714867743ddf5911ee7dff2717ae275f870551d86469cbdbf026c6b552e) + optionalmmproj(907,542,880 B,SHA256 = ef2af1d6aa8a22740c7252bc7ce40c5c2f1249aa8d9c8499a4d4572c3cfdf9df).MaxQ-32GB-Abliterated: Shard 1 =26,821,627,808 B(1,127tensors:32edited [27weight tensors +5blk.31..35F32ffn_down_exps.biasaligned with signed layerbeta_m = 0.449..1.710],1,095bit-identical toMaxQ-32GB;SHA256 = 09b54c335505f9f609415dbc774868ad5ae188f17e873dead3ec743ec84e1009) + Shared Shard 2 (86b79714...) + optionalmmproj(ef2af1d6...).
- Gated-Residual Over-Projection,
blk.48MTP Synchronization & Edition Positioning: BecauseQwen3.8-Flash-Nextroutes activations through widened multi-stream Gated Residual Hyper-Connections (hc_attn_*,hc_ffn_*), single-branchbeta <= 1.0projection leaves parallel residual branches carrying refusal state at late layers (blk.32..47).v3.0-r2encodesHeretic-2's TPE-optimized over-projection (0 / 100refusals vs.99 / 100upstream at0.0818KLD) plus signedbeta = 2.2508blk.48NextNMTP synchronization (81–85%draft-mtpacceptance on live multi-turn tool-calling at13.2–14.8 t/s,VmHWM = 46.34 GiBon64GBand28.40 GiB <= 28.5 GiBon32GB), positioning AlignedMaxQ-64GB/MaxQ-32GBas the default pure software-engineering flagships and*-ablitstrictly for opt-in security research, reverse engineering, red-team auditing, and uncensored legal/policy analysis (configure--temp 0.6 --top-p 0.95 --top-k 20 --presence-penalty 0.2 --repeat-penalty 1.05on32GBpresets to prevent low-bit greedy tail loops).
- Applies TPE-optimized signed multi-layer over-projection (
- 16 GB
NVIDIA-Nemotron-3.5-30B-A3BDonor-FreeMaxQ-16GB+ Zero-Recompute42/43 -> 44/45Fleet Unification:NVIDIA-Nemotron-3.5-30B-A3B-MaxQ-16GB(14,382,336,640 B=13.3946 GiB < 13.50 GiB,SHA256 = 26b884d70a52046a7198c4e01cac397af8bf44584b04cdcf59f896145587db67): Built 100% donor-free fromnvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16(401standardggmltensors:{F32: 237, Q4_0: 70, Q8_0: 48, MXFP4(39): 28, Q2_0(42): 18},max_type = 42,0custom types;0.000000%8/13-gram calibration overlap vswiki-alt.txt) for native unmodifiedllama.cpp, Apple Silicon Metal, CUDA, Vulkan, andLM Studio. Certified under hardware-enforced Linux Cgroup v2 (MemoryMax = 15.0G,MemorySwapMax = 0,-c 32768 -ub 512 -fa on -ctk q8_0 -ctv q4_0,oom = 0) at/proc/<pid>/statusVmHWM = 14.3040 GiB <= 15.0 GiB(RssAnon = 0.9097 GiB+RssFile = 13.3942 GiB; processVmHWMquoted over cgroupmemory.peak = 0.949 GiBARC/page-cache delta), delivering188.29 t/smean (194.05peak)pp512and15.10 t/smean (15.75peak)tg128pureN=1(3/3bit-exact) plus27.80 t/s(94.44%accept) onMiniLRUcode refactoring (merge_intervals+ OpenAI multi-turn tool-callingPASS).- Zero-Recompute
42/43 -> 44/45Migration & Transparent40 × 512wiki-alt.txtLadder (vs.16.07 GiBMXFP4ref6.5962 ± 0.1729): Patches strictly47header bytes (46ttype+1general.file_type,0payload bytes altered) acrossVQ3-64(14,282,559,040 B,SHA256 = 04a460cdd831002d80b2f47ec5d70e7a92227abcf0a57f31ad73fe0bd50e711f,PPL = 6.7614,Mean KLD = 0.0871— highest-accuracy16 GBCPU-Turbo flagship),Fractal-64(15,420,034,624 B,SHA256 = 3a3a09eac29b5906397ca6a93d2ff0f566fca0bdd365294d067aeb5d0ebf4a1b,PPL = 6.6533,Mean KLD = 0.0688), andAbliterated-VQ3-64(14,282,559,488 B,SHA256 = 75da8372f3896bfacd103f3cb85bf22b729d39cb1efce82fb3c838ab2af53e57), alongside universal std-ggmlMaxQ-16GB(PPL = 7.3129 ± 0.1923,Mean KLD = 0.1722,Median KLD = 0.0698,Same top p = 83.431%) and 64 GBDeepSeek-V4-Flash-0731-MaxQ-v4(57.99 GiB,6112e887.../5d85c740.../a8654d4d...). (Note: Nemotron's default Jinja template enables<think>...</think>reasoning; setmax_tokens >= 4096for thinking or passenable_thinking: false/--reasoning-budget 0for concise direct/JSON replies.)
- Unified
llama-modchip v3.0Engine Patch (patches/llama-modchip-v3.0.patch) & Zero-Collision Custom Types (44&45):- Zero Upstream Type Collision: Preserves upstream
GGML_TYPE_Q2_0 = 42(18 B / 64 W=2.25 bpw,QK2_0 = 64) andGGML_TYPE_F8_E4M3 = 43untouched; relocatesFractal-64(30 B / 64 W) toGGML_TYPE_FRACTAL_64 = 44andLUT-VQ3-64(24 B / 64 W) toGGML_TYPE_VQ3_64 = 45. A single binary built fromggml-org/llama.cpp @ de32b268c4ad55d45f3b7b16c029f4e077f4ba74+llama-modchip-v3.0.patch(SHA256 = b14475aa5350d632551b5f4261017e5ccb648750f8c6c9975255a171c367fbf0) runs the entire16 GB / 32 GB / 64 GBfleet across Linux, macOS Metal, Windows, CUDA, and Vulkan. - Zero-Copy 3D MoE
--repackCure &q8_cache/params->wdataAliasing Fixes: Fixes upstreamggml_compute_forward_mul_mat_tiled/mul_mat_id_tiledq8_cacheworkspace aliasing (ggml_cpu_get_q8_wdata), clearslast_mmid_ids_dataon every non-MUL_MAT_IDnode, fixesrepack.cppstride/unaligned-load bugs, skips heap repacking on 3D MoE tensors (ne[2] >= 64) so zero-copymmapis preserved (MaxQ-64GBRssAnonreduced5.4×from9.81 GiBto1.81 GiB,draft-mtpacceptance restored from0%to93.00%,batch=1--repackvs--no-repackKLD = 0.00000 ± 0.00000), and alignsMADV_HUGEPAGEinward to2 MiBboundaries (src/llama-mmap.cpp) so lazy PLE ranges are never pre-faulted. - 512-Bit
Q1_0/Q2_0/MXFP4Kernels & Multi-ThreadedGATED_DELTA_NET: Adds AVX-512 VNNI/BW single-col & multi-col (1x8/4/3/2) kernels forQ1_0 (41), signed-scaleQ2_0 (42), overflow-freeQ8_0, andMXFP4, eliminates inner barriers in multi-threaded 8-zmmFMAGATED_DELTA_NET, restores4*nthdynamic causal-triangle load balancing onFLASH_ATTN_EXTprefill (neq1 > 1), and widensStockfish-PVS v2/RouteCracktoken IDs to 32-bitllama_token(248,320vocab safe).
- Zero Upstream Type Collision: Preserves upstream
- Empirical Speedups, Transparent Attribution, & Runtime Defaults (
modchip,install.sh,scripts/llama_modchip.py):- Combined
v3.0Engine + Non-Convoy-t $((N_CORES - 1)) --poll 0Throughput (16Zen 5 vCPUs, default--repackON):MaxQ-64GB(draft-mtp):pp51239.24 ->71.97 t/s(+83.4%),tg1287.329 ->12.662 t/s(+72.8%vs Stage 4C--no-repack,4.95×vs broken pre-v3.0repack, peak12.82 t/s,1.988×overv3.0 none6.368 t/s,93.00%MTP accept,VmHWM = 45.08 GiB).MaxQ-32GB(draft-mtp):pp51261.39 ->74.55 t/s(+21.5%),tg1289.455 ->11.454 t/s(+21.1%onp512,14.29 t/savg /14.77 t/speak oninplace_edit,1.718×overv3.0 none6.668 t/s,VmHWM = 26.91 GiB <= 28.5 GiB).- (Transparent Attribution Note: Stage 4C ran oversubscribed
-t 16 --poll 50 --no-repackon 16 vCPUs, whereas Stage 4D-1 ran-t 15 -tb 15 --poll 0 --poll-batch 0withv3.0default--repackON; isolated non-convoy threading accounted for+25.9% pp / +31.3% tgon32GB none, and./modchip+install.shnow defaultllama-serverto-t $((N_CORES - 1)) -tb $((N_CORES - 1)) --poll 0 --poll-batch 0.)
248KSpeculative Shootout &RC_TABLES_DIR=/nonexistentIsolation: Built-indraft-mtpwon cold single-turnrep 0tg128on6 / 6(100%) of(edition, prompt)pairs; zero-weightStockfish-PVS v2(--spec-type ngram-simple --spec-ngram-simple-size-n 3 --spec-ngram-simple-size-m 3 --spec-ngram-simple-min-hits 1) achieved10.58–12.82 t/s(1.67×–1.87×) on coldrep 0source-in-context refactoring with+5.6%to+11.0%faster prefill and-0.53to-0.85 GiBlowerVmHWM../modchipandstart_server.shautomatically exportRC_TABLES_DIR=/nonexistentwhenever serving non-Nemotron models so./workspace/routecrackNemotron tables are never loaded by mistake./dev/shmHybrid State Save/Restore (MaxQ-32GB,512tok): Restores the233.14 MiB(244,461,872 B)36 GDN + 12 QSAstate in38.49 msavg (38.16 msmin,6.11 GiB/s,161.96×faster than6.18sprefill;~19 ms/120.6 MiBwith--ctx-checkpoints 0) with 100% bit-exact continuation (SHA256 = 8e5a6444...).- Sampling &
<think>Guidance:MaxQ-32GB(1.63 bpw:2/8blind pass at greedytemp=0,3/8greedy repetition loops without penalties vs.MaxQ-64GB5/8pass [5/5within4,096cap,0/8loops]) defaults to--temp 0.6 --top-p 0.95 --top-k 20 --presence-penalty 0.2 --repeat-penalty 1.05in./modchip; when enabling<think>mode, setmax_tokens >= 16384(orenable_thinking=Falsefor<= 4096caps). Unpatched upstreamllama.cppusers should pass--no-repack(-nr); not needed onv3.0.
- Combined
- 12 Static-
libgompPortableGLIBC_2.30Linuxx86_64Binaries (SHA256SUMS):- Links
libgomp.a,libstdc++.a, andlibgcc.astatically (NEEDED = [libdl.so.2, libpthread.so.0, libm.so.6, libc.so.6, ld-linux-x86-64.so.2],max_glibc <= 2.30,0external OpenMP runtime needed) and removes internal weight-building scripts from the public runtime tree:54af3232375ede820f91143c106a594c83b815a3dca66fe98ab08bcb8b63d4f0 llama-server-modchip-linux-x86_64-avx5120f0c07f97a98d8aeb9a012264fdeff5a787d2104a2afb122052bfe0a21c2d1a9 llama-server-modchip-linux-x86_64-avx2(andllama-server-modchip-linux-x86_64)531d870988dcc580322341d2bbd7f0cf6b6cc3b3376a2e86d092a021f7cddba2 llama-cli-modchip-linux-x86_64-avx51249a918206afe371d16ab14fd66d9cc87e35c7f8c0471c2cc13c759f2e0350d03 llama-cli-modchip-linux-x86_64-avx2(andllama-cli-modchip-linux-x86_64)7ad8253fb9c3a779be58122adc51ea9786bb391acb49d3cbfa6280cedb01e9d8 llama-bench-modchip-linux-x86_64-avx512fa8a055799bb2f65df8c6077f1377b81bfeda4efe3851639d05c1b44c3fa45e8 llama-bench-modchip-linux-x86_64-avx2003f452008972fb8696f7ea2800cfede360ba2a2c32f0a605fba2463b66bcadb llama-perplexity-modchip-linux-x86_64-avx512f77c3581d37fa6920daec6938abc5767675b430130eed03ec05e68c6e6d4b624 llama-perplexity-modchip-linux-x86_64-avx2befa4842fa09a061d4fe3376d226ed4ad2dafa8c149db10c06b37f85e59e3f94 llama-quantize-modchip-linux-x86_64-avx512521d9bf53f888c71e158732cd2c0b14e80a5e3d2005aed5f7a3e204ebfba9bbb llama-quantize-modchip-linux-x86_64-avx2
- Links
Official MaxQ v3.0 Hugging Face Lineup & Collection
- Official Collection (
MaxQ v3.0): huggingface.co/collections/Synchopate/maxq-v30-6afb6e2d63dfd573b4b1a95f - 64 GB Tier Flagship:
Synchopate/Qwen3.8-Flash-Next-MaxQ-64GB-GGUF(+ Opt-InMaxQ-64GB-Abliterated) &Synchopate/DeepSeek-V4-Flash-0731-MaxQ-v4-64GB-CPU-GGUF - 32 GB Tier Flagship:
Synchopate/Qwen3.8-Flash-Next-MaxQ-32GB-GGUF(+ Opt-InMaxQ-32GB-Abliterated) - 16 GB Tier:
Synchopate/NVIDIA-Nemotron-3.5-30B-A3B-VQ3-64-CPU-GGUF(v3.0CPU-Turbo) &Synchopate/NVIDIA-Nemotron-3.5-30B-A3B-MaxQ-16GB-GGUF(100%Standardggml— CPU / Metal / CUDA / Vulkan / LM Studio)