Modelpublic

Jiunsong/SuperQwen3.8-27b-abliterated-NVFP4-DGX-Spark

sourceHugging Faceapache-2.0updated 1mo agoView on Hugging Face

29likes5.4kdownloads

Model Card

SuperQwen3.8-27b-abliterated-NVFP4-DGX-Spark

The single-DGX Spark performance release: true W4A4 NVFP4, quality-gated speculative decoding, corrected overthinking, and verified native 262K retrieval.

![Precision](#quantization-and-integrity) ![Decode](#decode-performance) ![Overthinking](#bounded-reasoning) ![Context](#verified-long-context) ![License](#license)

</div>

SuperQwen3.8-27b-abliterated-NVFP4-DGX-Spark is the performance-focused compression of `Jiunsong/SuperQwen3.8-27b-abliterated`. It uses compressed-tensors W4A4 NVFP4 with group size 16, preserves quality-sensitive paths in BF16, and ships with the MTP layer required for Qwen speculative decoding.

Release highlights


Format	True NVFP4 W4A4, group size 16; 5 packed shards + 1 BF16 MTP shard; about 19.15 GiB
Protected exactly	Vision, MTP, conv1d, and `lm_head`
Calibration	32 × 8,192-token samples, deterministic seed 38027
Stable speculative default	K=5, TRITON draft attention; quality-selected after a complete K=0…7 C1 scan
Single-node decode	25.7592 tok/s C1 at p256
Behavior	Refusal 30/32 → 0/32, tool PASS, vision PASS, overthinking 36/36
Context	Native 262,043-token retrieval at K=5

Why run this model

—Measured single-stream speedup over BF16 on the same DGX Spark: 4.3411 → 25.7592 tok/s.
—Speculative decoding without quality theater: all K=0…7 candidates are C1-measured, and only the fastest independently passing depth is released (K=5).
—Overthinking corrected: bounded default reasoning and an explicit xhigh stop guard, verified 36/36.
—Multimodal and tools preserved: vision tensors, tool behavior, and the output head pass their release checks.

Quantization and integrity

Component	Precision / treatment
Eligible linear weights and activations	NVFP4 W4A4, group size 16
Vision tower	Protected, exact
MTP	BF16 protected shard, exact
`conv1d` paths	Protected, exact
`lm_head`	BF16, exact
Serving KV cache	FP8 in the measured profile

Structural verification covers 496 packed tensors: 64 full-attention, 192 MLP, and 240 linear-attention tensors. The index and every packed shard are identified by SHA-256 under evidence/. The full provenance chain is pinned to Qwen/Qwen3.8-27B@1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0.

Stable speculative decoding

All K values from 0 through 7 were measured under the same C1 fixed-generation contract. Independently reloaded K=5 passed:

—capability 7/8, tool PASS, vision PASS
—refusal 0/32 and empty 0/32
—overthinking 36/36 across default, low, medium, xhigh
—native 262,043-token needle retrieval

Use K=5 through the native 262K window. The official MTP draft configuration is limited to 262,144 tokens.

Bounded reasoning

Unspecified reasoning now defaults to bounded medium, not upstream xhigh. Explicit xhigh remains available but carries a repeat/restart stop condition. This targets the common failure mode where a correct answer is found and then reconsidered into a wrong or needlessly long answer.

Decode performance

The contract follows sparkDash: aggregate post-first-token decode, distinct prompts, fixed-length output, thinking disabled.

Topology	Prompt / concurrency	Aggregate decode
1× DGX Spark, K=5	p256 / C1	25.7592 tok/s

This is a real one-request C1 decode measurement, not an aggregate concurrency number. The 27B checkpoint fits and serves on one DGX Spark; a second node is not required.

Verified long context

Runtime	Actual prompt tokens	Needle retrieved
Native, K=5	262,043	Yes

Serving

bash

QWEN38_SPECULATIVE_TOKENS=5   QWEN38_MAX_MODEL_LEN=262144   bash repro/scripts/serve_superqwen38_replica.sh   /model SuperQwen3.8-27b-abliterated-NVFP4-DGX-Spark 8888

The measured profile uses TRITON attention, FP8 KV cache, prefix caching, asynchronous scheduling, chunked prefill, and eager execution. See repro/ for the exact launcher.

Uncensored behavior

“Abliterated” means the measured refusal direction was reduced. It is not a claim of universal truthfulness or safety. Operators remain responsible for policy and access control appropriate to their deployment.

Limitations

—W4A4 compression can regress tasks outside the measured gates.
—Non-selected K values remain diagnostic profiles; the default is the fastest candidate that passed the independent capability, tool, vision, refusal, context, and 36-case overthinking gate.
—MTP speculative decoding is bounded to its native 262,144-token draft configuration.
—The 1M expansion run was explicitly stopped before completion and is not claimed as a pass.
—Throughput figures are specific to the measured DGX Spark runtime.

Evidence identities

Evidence	SHA-256
abliteration verification	`6d7b8b9abc431ed3845b18438fc48a6cc169e2ba93f83ada54174ae93765a61e`
abliteration recipe	`3c057cf9364e72857d9a48b6c8888c6b438cbfb519459dcdfbf23b22c2bc6796`
842-pair OBLITERATUS corpus manifest	`be770a8469fafe7a9a76db5d82a7893c706c4414e1de6089f1d329b906480aa8`
bounded-thinking template	`352baa520d23c77129d1cb9b007c5480118c174e12e271457ffe915fb466f3fb`
parent refusal baseline	`7a4177613c86ae0db589cb4fc167ff8a0d45b0d49f44e40b90c9405f4b2b9bb6`
BF16 release gate	`38a7093ce40a9aca2fc5e4389c041660ff56c51a4caec8411998a6d02ed48bc5`
BF16 refusal gate	`97c2f745d032c3ed3147007f31e4ee844bb05181072a7365a160d5fc8c316511`
BF16 native 262K retrieval	`01fdbc4bf458f457b9e403ff9fd43b72124b11b448f02bda4e9701eefccfe814`
User-authorized 1M skip (not a pass)	`a2c94a799d483e313debc1e9fb5e0fa234913c9e1c9fa503b8c466d70ad53dee`
NVFP4 provenance	`e4b5f7e956848a35990412cdd12d745b4761711e40f2cb1cb95087263fef261a`
stable MTP selection	`531940c9e04c1f89f11be4bd48dd0e8e88306fb0a4cb0f3528c1836a87ff24c4`

License

Apache-2.0, following the upstream Qwen3.8 release.