license: apache-2.0
library_name: transformers
pipeline_tag: image-text-to-text
base_model: Qwen/Qwen3.8-27B
base_model_relation: finetune
tags:
- qwen3.8
- qwen3.5
- multimodal
- image-text-to-text
- reasoning
- tool-calling
- long-context
- 1m-context
- uncensored
- abliterated
- obliteratus
- supertune
- vllm
- nvfp4
- w4a4
- quantized
- dgx-spark
- speculative-decoding
language: - en
- ko
SuperQwen3.8-27b-abliterated-NVFP4-2xDGX
The DGX Spark performance release: true W4A4 NVFP4, stable K=1 speculative decoding, corrected overthinking, two-node scale-out, and verified 1M retrieval.
SuperQwen3.8-27b-abliterated-NVFP4-2xDGX is the performance-focused compression ofJiunsong/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=1, TRITON draft attention; quality-selected after a complete K=0…7 scan |
| Single-node decode | 17.7532 tok/s C1, 86.1382 tok/s C6 aggregate at p256 |
| Two-node scale-out | 95.9850 tok/s C6 aggregate across two independent DGX Spark replicas |
| Behavior | Refusal 18/32 → 0/32, tool PASS, vision PASS, overthinking 36/36 |
| Context | Native 262,043-token retrieval at K=1; 1,000,045-token YaRN retrieval at K=0 |
Why run this model
- About 2.4× faster than BF16 at C6 on the measured single DGX: 38.5630 → 86.1382 tok/s.
- Two-node aggregate throughput: 95.9850 tok/s with independent replicas and round-robin dispatch.
- Speculative decoding without quality theater: raw K=3/K=6 speed winners were rejected when quality regressed; K=1 passed every release gate.
- 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 toQwen/Qwen3.8-27B@1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0.
Stable speculative decoding
All K values from 0 through 7 were measured. K=3 was the raw C6 speed winner and K=6
the raw C1 winner, but K=3/K=4 reproduced a deterministic translation overthinking
failure. They were rejected. Independently reloaded K=1 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=1 through the native 262K window. Use K=0 for the verified YaRN 1M
profile, because the official MTP draft configuration itself is limited to 262,144.
Bounded reasoning
Unspecified reasoning now defaults to bounded medium, not upstream xhigh. Explicitxhigh 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=1 | p256 / C1 | 17.7532 tok/s |
| 1× DGX Spark, K=1 | p256 / C6 | 86.1382 tok/s |
| 2× independent DGX Spark replicas, K=1 | p256 / C6 | 95.9850 tok/s |
These are aggregate concurrent figures. They are not single-stream speed, and the
two-replica result is not mislabeled as TP=2.
Two-node DGX Spark
The recommended high-throughput topology is one K=1 replica per DGX Spark with
round-robin request dispatch. It avoids slow cross-node collectives and scales complete
requests independently.
- Both nodes run the same artifact bytes and the same serving flags.
- C6 uses six distinct concurrent prompts, three routed to each endpoint.
- K=1 is the release default; K=0 is the exact-quality and 1M fallback.
Verified long context
| Runtime | Actual prompt tokens | Needle retrieved |
|---|---|---|
| Native, K=1 | 262,043 | Yes |
| YaRN, K=0 | 1,000,045 | Yes |
Serving
QWEN38_SPECULATIVE_TOKENS=1 QWEN38_MAX_MODEL_LEN=262144 bash repro/scripts/serve_superqwen38_replica.sh /model SuperQwen3.8-27b-abliterated-NVFP4-2xDGX 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.
- K>1 was intentionally not released as the default because deterministic quality regressed.
- MTP speculative decoding is bounded to its native 262,144-token draft configuration.
- 1M is verified with YaRN and K=0; latency is substantially higher than native context.
- Throughput figures are specific to the measured DGX Spark runtime.
Evidence identities
| Evidence | SHA-256 |
|---|---|
| abliteration verification | 3355e3843b6f7765fdae7e7c41593132d75e311f7b5c6371a8afaadee7675734 |
| bounded-thinking template | 11eadfd51d237324252e887e65daf59c55f10fb0f1d0b71c9b8f131c88ba9d4e |
| parent refusal baseline | c1c8f8672f0a40a57c20153445fc7d76290b59716fc5ca9a6fb03881976c7a99 |
| BF16 release gate | 64639420cf9b0d2c5a25dd71b96a453092a26e600d74315257cc2633211c180b |
| BF16 refusal gate | cf9683dcaca9eb741d138b6de3c4d53b15f747d27c0071b0b9538716123c4903 |
| BF16 1M retrieval | 7e59e289df043104880c5e2569192c78ffe2aa0056e9c3c2b8244089b41dfc09 |
| NVFP4 provenance | 0da8a09e55b29bf6dee4c78d97a1cf5fac56f16e00ce02c7df3d9b21fe85b767 |
| stable MTP selection | a11b7b7841152c64a501a4aeb0dd7ac2cdbf5eb896ed35a13a8512eafcd334ff |
| NVFP4 1M retrieval | 78829b89b329c50296d67eb3729e7b79e483be1ecdf2b3cf2120beb758e6d4f5 |
| two-node replica throughput | 32ff903cb9ca0790b565ab45b6b501dc70ba0db09b17be647d7d235bc3fce363 |
| TP=2 comparison | d974ee3aad7d95185c71ad3f28dd35b9e2981e9f8682c4e6306b4ac2b4fbd847 |
License
Apache-2.0, following the upstream Qwen3.8 release.