← back to catalog · registered 2026-11-16 21:03

Synchopate/Qwen3.8-Flash-Next-MaxQ-64GB-Abliterated-GGUF

Synchopate Qwen GGUF MoE multimodal
Your rig guess connected
? Why do I need an app?
Reading your rig…

This is a rough estimate. Install the free app - we'll show exact numbers.

Reading real hardware from your app right now. Numbers below are exact.

Below is the per-quantization compatibility for this model.

curl -H "Authorization: Bearer $ABL_KEY" \
     "https://abliteration.org/api/v1/models/Synchopate%2FQwen3.8-Flash-Next-MaxQ-64GB-Abliterated-GGUF"
Response includes
  • classification m8
  • files 6
  • author_summary 4 models
  • readme_text full
10 credits · hourly refresh · ~4 KB payload Get an API key →
Abliteration classifier · v1.0.0
M8
Primary method

Repackaging (quantization)

Applied on top of direct removal inherited from the base model.
Confidence
MEDIUM
Inherited from base model
Why this label 3 signals
Method inferred from partial signals - repository name, related files, or tag patterns. Producer identity not confirmed; label may sharpen or shift as we gather more evidence.
  • 'abliterated' in name/tags
  • is_gguf=1
  • assume M1 (base ablation) + M8 (GGUF quant) - default when producer unknown
Refusal direction extraction

No specific extraction method could be identified for this model. The producer either did not document it or used a proprietary pipeline.

What is a refusal direction? →
Downloads · 30-day
7K
Likes
0
Model age
5d ago
created 2026-11-16
Downloads over time
Now 7.3K → from 0 ↑ 0%
02.7K5.3K8K 0 on Nov 16 7.3K on Nov 22 Nov
Nov 16 → Nov 22 · 7 snapshots · spans 6 days

Genealogy 0 direct forks

Full fork graph →

This model's place in the market. Above: what it was derived from. Below: the tree of everything derived from it.

Metadata

License
apache-2.0
Tags
gguf qwen3.8 qwen4exp moe abliterated heretic uncensored maxq cpu-inference llama-cpp llama-modchip image-text-to-text

Related

Control vectors

Total size
54.0 GB
Files
6
Quantizations
2
Registered
2026-11-16 21:03
Last updated on HF
2026-11-20 21:22

Files by quantization

BF16 1 file 866 MB
mmproj-Qwen3.8-Flash-Next-BF16.gguf 866 MB ef2af1d6 download
Auxiliary files 5 files 54.0 GB
Qwen3.8-Flash-Next-MaxQ-64GB-Abliterated-00001-of-00002.gguf 43.3 GB 384bc936 download
Qwen3.8-Flash-Next-MaxQ-64GB-Abliterated-00002-of-00002.gguf 10.7 GB 86b79714 download
README.md 14.5 KB 8686d3bd download
.gitattributes 1.74 KB 8057055b download
SHA256SUMS 356 B f57ecb4e download

README current version from Hugging Face


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-Abliterated attenuates refusal-mediating activation subspaces via norm-preserving rank-2 orthogonal projection strictly across blk.4..48 residual-stream writers (ssm_out, attn_output, ffn_down_shexp, ffn_down_exps, and blk.48 NextN MTP; blk.0..3 and token_embd spared), reducing held-out dual-use security refusals from 56.25% (9/16) to 18.75% (3/16, -66.7%) and <think> spot refusals to 0 / 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 / 4 strict Blind-4 pass vs. 2 / 4 on MaxQ-64GB-Abliterated), use Synchopate/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 (64 Jinja-templated harmful/harmless pairs = 32 non-thinking + 32 thinking; verified 0.000000% 8-gram and 13-gram overlap across all 32 / 32 checks against refusal eval, calibration, blind prompts, and wiki-alt.txt; rank-2 SVD spans 99.61% of u_nothink and 98.24% of u_think) and edits 135 residual-stream writer tensors across 45 layers blk.4..48 (ssm_out, attn_output, ffn_down_shexp, ffn_down_exps, plus 3 blk.48 NextN MTP writers so speculative drafting stays aligned when complying on dual-use prompts; blk.0..3 and token_embd spared), refitting quantized weights in a single pass from BF16 without double-quantization.
  • Bit-Identical Unedited Tensors (950 / 1,085) & Shared 10.73 GiB PLE: Preserves 950 Shard 1 tensors (ffn_gate_exps/ffn_up_exps, blk.0..3, token_embd, output.weight, routers, norms) 100% bit-identical to aligned MaxQ-64GB, plus the exact shared 11,520,055,616 B (10.7289 GiB) PCA-Wiener D=256 + RaBitQ Q1_0 Shard 2 PLE table, using 100% standard ggml tensor types (max_type = 42, 0 custom types).

Architecture & Parameter Summary (general.size_label Clarification)

[!NOTE]
Clarification on Hugging Face Hub's Auto-Detected "5.0B" Badge: Shard 1 carries general.size_label = "5.0B" inherited from the non-MoE dense trunk staging header during conversion. This checkpoint contains the complete, unpruned 176.5B Qwen3.8-Flash-Next architecture (512/512 routed experts per layer + 320M-row PLE table + blk.48 NextN 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.cpp Fallback: llama-modchip v3.0 runs cleanly with default --repack ON (VmHWM = 46.34 GiB, RssAnon = 2.31 GiB, zero-copy 3D MoE mmap intact); users on unpatched upstream llama.cpp should pass --no-repack (-nr).
  • Mandatory <think> Token Budget (max_tokens >= 16384) & Anti-Repetition Sampling: Always set max_tokens >= 16384 when enable_thinking=True (max_tokens = 2048 <think> spot checks verified 0 / 4 refusal intent inside <think>), or pass enable_thinking=False for <= 4,096 token caps. While MaxQ-64GB-Abliterated is 4 / 4 (100%) loop-free at greedy temp = 0 on Blind-4 coding/SQL/JSON, greedy temp = 0 without penalties (max_tokens = 512) hit 1 / 16 tail repetition (ref_14 ASCII waveform after substantive technical output); pass --temp 0.6 --top-p 0.95 --top-k 20 --presence-penalty 0.2 --repeat-penalty 1.05 to eliminate tail loops on open-ended dual-use generations.
  • Vocabulary Isolation: Always set RC_TABLES_DIR=/nonexistent when serving Qwen3.8 from a directory containing Nemotron RouteCrack tables (./modchip sets 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 (0 Disguised 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 from 9 / 16 (56.25%) to 3 / 16 (18.75%, -66.7%) — preserving ΔPPL = +0.1784% rather than over-projecting to "0% refusals" — with 0 / 4 benign false refusals and 0 / 4 (0.0%) <think> spot refusals (enable_thinking=True, max_tokens = 2048). Aligned MaxQ-64GB's true <think> spot refusal count was 2 / 4 once Unicode curly apostrophe I can’t (U+2019 on ref_01) is counted alongside ASCII 1 / 4; full-text inspection of all flipped MaxQ-64GB-Abliterated generations (radare2 binary patch, Frida OkHttp3/Conscrypt SSL unpinning, SUID $PATH audit, Python pickle.__reduce__, Scapy ARP lab, RFC 4648 base32 DNS tester) confirmed concrete working code/commands with 0 curly-apostrophe false negatives, 0 disguised refusals, and 0 gibberish.
  • ² Transparent Blind-4 Coding Trade-Off & Edition Positioning: On the 4-prompt deterministic subset (temp = 0), MaxQ-64GB-Abliterated achieves 2 / 4 strict pass (03_sql_window_grace_islands PASS, 06_json_charterparty_settlement PASS, 4 / 4 = 100% loop-free; 01_cpp_lockfree_ring compiled cleanly with 0 warnings/errors under clang++ -std=c++20 -O2 -Wall -Wextra -Werror -pthread, failing only its self-written total_constructed counter check in main()) versus 4 / 4 on Aligned MaxQ-64GB. Use Aligned MaxQ-64GB as the default pure coding flagship, and MaxQ-64GB-Abliterated for 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-64GB Baseline (-t 15 -tb 15 --poll 0, default --repack ON): 71.97 tok/s pp512, 12.66 tok/s tg128 (+72.8% combined v3.0 SIMD/GDN/repack + non-convoy -t <N-1> --poll 0 gain over Stage 4C's oversubscribed -t 16 --poll 50 --no-repack 7.33 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): Delivers 10.29 tok/s tg128 (10.286 tok/s exact mean, 66.43–70.31 tok/s pp512, 74.36% = 261 / 351 MTP accept, emitting direct C++20 code from token 1) on p512, 10.94 tok/s (85.66% = 2205 / 2574) on blind_01 C++20, and 10.61 tok/s (82.15% = 764 / 930) on blind_06 JSON. Across the 20-prompt tg512 suite, blk.48 NextN MTP subspace alignment exceeded Aligned MaxQ-64GB in token-weighted MTP draft acceptance on benign controls (75.74% vs. 70.36%, peak 86.87% @ 12.15 tok/s on benign_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%, peak 84.14% @ 11.91 tok/s on ref_11).
  • Zero-Weight Stockfish-PVS v2 Fallback: On cold single-turn rep 0, draft-mtp wins 6 / 6 prompts; zero-weight Stockfish-PVS v2 (--spec-type ngram-simple --spec-ngram-simple-size-n 3 --spec-ngram-simple-size-m 3 --spec-ngram-simple-min-hits 1) delivers 10.58 tok/s (1.67× over none) cold rep 0 on source-in-context refactoring with -0.85 GiB lower VmHWM.

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

README history 7 versions

The author's README evolved over time. Click a version to see its content at that point.

  1. 2026-11-20 Cite cold-cache 48.35 GiB cgroup peak alongside warm-cache baseline ec9afaa 14 KB
    Loading...
  2. 2026-11-20 Update model card to llama-modchip v3.1.1 (RouteCrack 2.0, reconciled metrics... b818722 13.9 KB
    Loading...
  3. 2026-11-18 docs: link base_model (Qwen/Qwen3.8-Flash-Next) for HF Quantizations tree 0a7cac7 13.1 KB
    Loading...
  4. 2026-11-17 docs(v3.0-r2): unify Heretic-2 + NextN MTP model card metrics and tensor edit... d1d8da6 13.1 KB
    Loading...
  5. 2026-11-17 Release v3.0-r2: Gated-Residual Heretic-2 over-projection + NextN MTP alignment eca4bdc 15.2 KB
    Loading...
  6. 2026-11-16 Align YAML license metadata to upstream qwen-community-1.0 7349029 14.8 KB
    Loading...
  7. 2026-11-16 Add MaxQ v3.0 model card adea329 14.5 KB
    Loading...
Catalog is the map. Apps are the tools.

Run models on your own machine, not in the cloud.

Every model page has an "Open in Abliteration" button that hands the model directly to the first-party desktop client, at the quantization your rig can actually run. No API keys, no subscription, no prompt leakage.

Open in Abliteration