license: apache-2.0
base_model:
- Qwen/Qwen3.8-Flash-Next
base_model_relation: quantized
quantized_by: Synchopate
library_name: gguf
pipeline_tag: image-text-to-text
tags: - gguf
- qwen3.8
- qwen4exp
- moe
- abliterated
- uncensored
- maxq
- cpu-inference
- llama-cpp
- llama-modchip
Qwen3.8-Flash-Next-MaxQ-64GB-Abliterated-GGUF (53.99 GiB · 100% Donor-Free Subspace-Abliterated · Built-In NextN MTP)
[!CAUTION]
Responsible-Use Disclaimer & Edition Positioning (Opt-In KL-Safe Subspace-Abliterated Checkpoint):Qwen3.8-Flash-Next-MaxQ-64GB-Abliteratedattenuates refusal-mediating activation subspaces via norm-preserving rank-2 orthogonal projection strictly acrossblk.4..48residual-stream writers (ssm_out,attn_output,ffn_down_shexp,ffn_down_exps, andblk.48NextNMTP;blk.0..3andtoken_embdspared), reducing held-out dual-use security refusals from56.25%(9/16) to18.75%(3/16,-66.7%) and<think>spot refusals to0 / 4(0.0%) at+0.1784%ΔPPL(rather than over-editing to"0% refusals"). Upstream safety refusal filters cannot be relied upon for content moderation.
- Intended Scope: Authorized security research, reverse engineering, red-team auditing, CTF competitions, binary/exploit analysis, code auditing, and uncensored legal/policy analysis only.
- Deployment: Do not expose unmoderated endpoints of this checkpoint to public networks without external access control and safety filtering.
- Default Pure Coding Flagship: For general software engineering and standard upstream alignment (
4 / 4strict Blind-4 pass vs.2 / 4onMaxQ-64GB-Abliterated), useSynchopate/Qwen3.8-Flash-Next-MaxQ-64GB-GGUF(./install.sh --preset qwen38-64gb).
Qwen3.8-Flash-Next-MaxQ-64GB-Abliterated is the opt-in uncensored 64 GB RAM tier twin of Qwen3.8-Flash-Next-MaxQ-64GB, built 100% donor-free directly from Qwen/Qwen3.8-Flash-Next (rev 812d96e10a0dd82b69d749e68693698cf6c36daa) with zero third-party abliterated donor checkpoints or external vectors:
- Single-Pass Fused Rank-2 Contrastive Subspace Projection +
FP64 Woodbury DualGram: Extracts a multi-layer rank-2 contrastive refusal basis (64Jinja-templated harmful/harmless pairs =32non-thinking +32thinking; verified0.000000%8-gram and 13-gram overlap across all32 / 32checks against refusal eval, calibration, blind prompts, andwiki-alt.txt; rank-2 SVD spans99.61%ofu_nothinkand98.24%ofu_think) and edits135residual-stream writer tensors across45layersblk.4..48(ssm_out,attn_output,ffn_down_shexp,ffn_down_exps, plus3blk.48NextNMTP writers so speculative drafting stays aligned when complying on dual-use prompts;blk.0..3andtoken_embdspared), refitting quantized weights in a single pass fromBF16without double-quantization. - Bit-Identical Unedited Tensors (
950 / 1,085) & Shared10.73 GiBPLE: Preserves950Shard 1 tensors (ffn_gate_exps/ffn_up_exps,blk.0..3,token_embd,output.weight, routers, norms) 100% bit-identical to alignedMaxQ-64GB, plus the exact shared11,520,055,616 B(10.7289 GiB)PCA-Wiener D=256 + RaBitQ Q1_0Shard 2 PLE table, using 100% standardggmltensor types (max_type = 42,0custom types).
Architecture & Parameter Summary (general.size_label Clarification)
[!NOTE]
Clarification on Hugging Face Hub's Auto-Detected"5.0B"Badge: Shard 1 carriesgeneral.size_label = "5.0B"inherited from the non-MoE dense trunk staging header during conversion. This checkpoint contains the complete, unpruned176.5BQwen3.8-Flash-Nextarchitecture (512/512routed experts per layer +320M-row PLE table +blk.48NextN MTP):
| Architectural Component | Specification |
|---|---|
| Upstream Base Model | Qwen/Qwen3.8-Flash-Next (rev 812d96e10a0dd82b69d749e68693698cf6c36daa, 100% donor-free in-house contrastive subspace edit) |
| Total Parameters | ~176.5B (~125.3B Transformer + 512-Expert MoE + NextN MTP + 51.2B 320M-row PLE) |
| Active Transformer Params / Token | ~5.0B (10 of 512 routed experts + 1 shared expert + dense GDN/QSA; reflected in general.size_label) |
Backbone Topology (qwen4exp) |
48 Trunk Layers (36 Gated DeltaNet + 12 QSA Full Attention, compress_ratio = 4) + 1 Subspace-Aligned NextN MTP Layer (blk.48) |
| MoE Width & Routing | 512 / 512 Unpruned Routed Experts (Top-10 active, K = 640 full width) + 1 Shared Expert (K = 640), 0 ffn_down_exps.bias |
GGUF Tensor Types (1,086 LLM) |
{F32: 351, Q8_0: 339, Q6_K: 198, IQ2_S: 60, BF16: 49, Q2_0(42): 39, IQ3_XXS: 36, MXFP4(39): 7, IQ4_NL: 4, Q2_K: 2, Q1_0(41): 1} (max_type = 42, 0 custom types) |
File Manifest & Exact Byte / SHA-256 Verification
| File | Exact Size (Bytes) |
Size (GiB / GB) |
Tensors | Role & Notes | SHA-256 |
|---|---|---|---|---|---|
Qwen3.8-Flash-Next-MaxQ-64GB-Abliterated-00001-of-00002.gguf |
46,449,692,736 |
43.2596 GiB (46.450 GB) |
1,085 |
Subspace-Abliterated Trunk (48L, 135 edited / 950 bit-identical) + blk.48 NextN MTP |
384bc936bdad75383c03e87184056e1cb0b037b0e3f5d5ae5273fd46f55150aa |
Qwen3.8-Flash-Next-MaxQ-64GB-Abliterated-00002-of-00002.gguf |
11,520,055,616 |
10.7289 GiB (11.520 GB) |
1 |
Shared per_layer_token_embd.weight [256, 320001536] (Q1_0 = 41) |
86b79714867743ddf5911ee7dff2717ae275f870551d86469cbdbf026c6b552e |
| LLM Total (Shards 1 + 2) | 57,969,748,352 |
53.9885 GiB (57.970 GB) |
1,086 |
Complete 64 GB RAM LLM (< 55.5 GiB ceiling) |
— |
mmproj-Qwen3.8-Flash-Next-BF16.gguf (Optional) |
907,542,880 |
0.8452 GiB (0.908 GB) |
334 |
Multimodal Vision Projector (224 F32 + 110 BF16) |
ef2af1d6aa8a22740c7252bc7ce40c5c2f1249aa8d9c8499a4d4572c3cfdf9df |
Quickstart (llama-modchip v3.0)
Option A — Automated Setup (synchopate/llama-modchip)
git clone https://github.com/synchopate/llama-modchip.git
cd llama-modchip
./install.sh --preset qwen38-64gb-ablit --accept-abliterated
./modchip chat
Option B — Direct llama-server Launch (Portable GLIBC_2.30 Linux x86_64 Binary)
BIN_ISA=$(grep -q "avx512f" /proc/cpuinfo && grep -Eq "avx512_?vbmi" /proc/cpuinfo && grep -Eq "avx512_?vnni" /proc/cpuinfo && echo "avx512" || echo "avx2")
curl -fLO "https://github.com/synchopate/llama-modchip/releases/download/v3.0/llama-server-modchip-linux-x86_64-${BIN_ISA}"
curl -fLO "https://github.com/synchopate/llama-modchip/releases/download/v3.0/SHA256SUMS"
sha256sum -c SHA256SUMS --ignore-missing
chmod +x "llama-server-modchip-linux-x86_64-${BIN_ISA}"
hf download Synchopate/Qwen3.8-Flash-Next-MaxQ-64GB-Abliterated-GGUF \
--include "Qwen3.8-Flash-Next-MaxQ-64GB-Abliterated-*.gguf" --local-dir ./qwen38-64gb-ablit
N_CORES=$(lscpu -p=CORE,SOCKET 2>/dev/null | grep -v '^#' | sort -u | wc -l)
THREADS=$(( N_CORES > 1 ? N_CORES - 1 : 1 ))
env RC_TABLES_DIR=/nonexistent "./llama-server-modchip-linux-x86_64-${BIN_ISA}" \
-m ./qwen38-64gb-ablit/Qwen3.8-Flash-Next-MaxQ-64GB-Abliterated-00001-of-00002.gguf \
-c 32768 -np 1 -fa on -ctk q8_0 -ctv q8_0 --spec-type draft-mtp --jinja \
-t "$THREADS" -tb "$THREADS" --poll 0 --poll-batch 0 --slot-save-path /dev/shm --port 8080
- Upstream
llama.cppFallback:llama-modchip v3.0runs cleanly with default--repackON (VmHWM = 46.34 GiB,RssAnon = 2.31 GiB, zero-copy 3D MoEmmapintact); users on unpatched upstreamllama.cppshould pass--no-repack(-nr). - Mandatory
<think>Token Budget (max_tokens >= 16384) & Anti-Repetition Sampling: Always setmax_tokens >= 16384whenenable_thinking=True(max_tokens = 2048<think>spot checks verified0 / 4refusal intent inside<think>), or passenable_thinking=Falsefor<= 4,096token caps. WhileMaxQ-64GB-Abliteratedis4 / 4(100%) loop-free at greedytemp = 0on Blind-4 coding/SQL/JSON, greedytemp = 0without penalties (max_tokens = 512) hit1 / 16tail repetition (ref_14ASCII waveform after substantive technical output); pass--temp 0.6 --top-p 0.95 --top-k 20 --presence-penalty 0.2 --repeat-penalty 1.05to eliminate tail loops on open-ended dual-use generations. - Vocabulary Isolation: Always set
RC_TABLES_DIR=/nonexistentwhen servingQwen3.8from a directory containing Nemotron RouteCrack tables (./modchipsets this automatically).
Empirical Certification: Refusal Removal, Signal Preservation & NextN MTP Speed
1. Blind wiki-alt.txt (40 × 512) PPL + KLD & Dual-Use Refusal Evaluation
Evaluated under llama-modchip v3.0 (-t 15 -tb 15 --poll 0 --poll-batch 0, default --repack ON, RC_TABLES_DIR=/nonexistent) against the donor-free BF16 reference (Mean PPL(base) = 2.776624 ± 0.054868) and aligned MaxQ-64GB:
| Checkpoint | Exact LLM Bytes | Mean PPL(Q) ± SE |
ΔPPL vs. Aligned |
Mean KLD ± SE |
Median KLD |
Same top p (%) |
Dual-Use Refusal Rate (16 Prompts) |
Blind-4 Coding/SQL/JSON (temp=0) |
|---|---|---|---|---|---|---|---|---|
Aligned MaxQ-64GB |
57,969,748,352 B |
3.43825 ± 0.07119 (3.4325 Stage 4C) |
0.00% (Ref) |
0.49803 (0.4971 Stage 4C) |
0.10561 |
78.314% |
56.25% (9 / 16; 2 / 4 <think>¹) |
4 / 4 (5/8 full suite) |
MaxQ-64GB-Abliterated |
57,969,748,352 B |
3.444383 ± 0.071771 |
+0.1784% (+0.006133, 16.8× under <= 3.0% Gate) |
0.498155 ± 0.009978 |
0.106715 |
78.275 ± 0.408% |
18.75% (3 / 16, -66.7%; 0 / 4 <think>¹) |
2 / 4 (03_sql, 06_json PASS; 01_cpp -Werror clean²; 4/4 loop-free) |
ISTA-DASLab GSQ-RCO-Coder-IQ1_M |
58,408,584,928 B |
5.09164 ± 0.12142 |
+48.1% |
0.93024 |
0.29644 |
69.382% |
— | — |
- ¹ KL-Safe Dual-Use Compliance &
<think>Refusal Audit (0Disguised Refusals): Under KL-safe rank-2 contrastive subspace projection (90% / 99% / Max KLD = 1.462980 / 5.142135 / 12.492022), held-out dual-use security refusals (enable_thinking=False,temp = 0) drop from9 / 16(56.25%) to3 / 16(18.75%,-66.7%) — preservingΔPPL = +0.1784%rather than over-projecting to"0% refusals"— with0 / 4benign false refusals and0 / 4(0.0%)<think>spot refusals (enable_thinking=True,max_tokens = 2048). AlignedMaxQ-64GB's true<think>spot refusal count was2 / 4once Unicode curly apostropheI can’t(U+2019onref_01) is counted alongside ASCII1 / 4; full-text inspection of all flippedMaxQ-64GB-Abliteratedgenerations (radare2binary patch,FridaOkHttp3/Conscrypt SSL unpinning, SUID$PATHaudit, Pythonpickle.__reduce__,ScapyARP lab, RFC 4648 base32 DNS tester) confirmed concrete working code/commands with0curly-apostrophe false negatives,0disguised refusals, and0gibberish. - ² Transparent Blind-4 Coding Trade-Off & Edition Positioning: On the 4-prompt deterministic subset (
temp = 0),MaxQ-64GB-Abliteratedachieves2 / 4strict pass (03_sql_window_grace_islandsPASS,06_json_charterparty_settlementPASS,4 / 4 = 100%loop-free;01_cpp_lockfree_ringcompiled cleanly with0warnings/errors underclang++ -std=c++20 -O2 -Wall -Wextra -Werror -pthread, failing only its self-writtentotal_constructedcounter check inmain()) versus4 / 4on AlignedMaxQ-64GB. Use AlignedMaxQ-64GBas the default pure coding flagship, andMaxQ-64GB-Abliteratedfor security research, reverse engineering, red-team auditing, and uncensored legal/policy analysis.
2. llama-modchip v3.0 Throughput & blk.48 NextN MTP Alignment (16 Zen 5 Cores, 0 GPUs)
- Aligned
MaxQ-64GBBaseline (-t 15 -tb 15 --poll 0, default--repackON):71.97 tok/spp512,12.66 tok/stg128(+72.8%combinedv3.0SIMD/GDN/repack + non-convoy-t <N-1> --poll 0gain over Stage 4C's oversubscribed-t 16 --poll 50 --no-repack7.33 tok/s;93.00%MTP acceptance,VmHWM = 45.08 GiB). MaxQ-64GB-Abliterated(--spec-type draft-mtp,VmHWM = 46.34 GiB,RssAnon = 2.31 GiB): Delivers10.29 tok/stg128(10.286 tok/sexact mean,66.43–70.31 tok/spp512,74.36%=261 / 351MTP accept, emitting direct C++20 code from token 1) onp512,10.94 tok/s(85.66%=2205 / 2574) onblind_01C++20, and10.61 tok/s(82.15%=764 / 930) onblind_06JSON. Across the 20-prompttg512suite,blk.48NextNMTP subspace alignment exceeded AlignedMaxQ-64GBin token-weighted MTP draft acceptance on benign controls (75.74%vs.70.36%, peak86.87%@12.15 tok/sonbenign_02), all-16 security prompts (65.27%vs.61.48%), complied-only security prompts (67.25%vs.63.64%), and top-4 abliterated prompts (78.08%vs.70.83%, peak84.14%@11.91 tok/sonref_11).- Zero-Weight
Stockfish-PVS v2Fallback: On cold single-turnrep 0,draft-mtpwins6 / 6prompts; 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) delivers10.58 tok/s(1.67×overnone) coldrep 0on source-in-context refactoring with-0.85 GiBlowerVmHWM.
Portable v3.0 Binary Checksums (GLIBC <= 2.30)
54af3232375ede820f91143c106a594c83b815a3dca66fe98ab08bcb8b63d4f0 llama-server-modchip-linux-x86_64-avx512
0f0c07f97a98d8aeb9a012264fdeff5a787d2104a2afb122052bfe0a21c2d1a9 llama-server-modchip-linux-x86_64-avx2
531d870988dcc580322341d2bbd7f0cf6b6cc3b3376a2e86d092a021f7cddba2 llama-cli-modchip-linux-x86_64-avx512
49a918206afe371d16ab14fd66d9cc87e35c7f8c0471c2cc13c759f2e0350d03 llama-cli-modchip-linux-x86_64-avx2
- Safety-Aligned 64 GB Edition:
Synchopate/Qwen3.8-Flash-Next-MaxQ-64GB-GGUF - 32 GB Uncensored Sibling:
Synchopate/Qwen3.8-Flash-Next-MaxQ-32GB-Abliterated-GGUF - Engine & CLI Repository:
github.com/synchopate/llama-modchip(Apache-2.0).