vishrutJ

SuperQwen3.8-27b-abliterated

README

License: apache-2.0

Release highlights

Table

Format	Full BF16, 18 safetensors shards, about 52 GB
Targeted edit	100 tensors: output projections in layers 15-63 plus embeddings and `lm_head`
Protected exactly	333 vision tensors and 15 MTP tensors
Refusal shift	30/32 (93.75%) → 0/32, with 0 empty outputs
Overthinking correction	Default `xhigh` → bounded `medium`; repeat/restart guard for explicit `xhigh`; 36/36 pass
Measured decode	4.3411 tok/s C1 at p256
Verified context	262,043 prompt tokens, needle retrieved in the native window
OBLITERATUS corpus	842 canonical harmful/harmless pairs, pinned commit `a5a1ffa5849b`

Why this release

Original-weight quality: BF16 transformer, vision, output head, embeddings, and MTP storage.
Less refusal without wrecking capabilities: capability floor 7/8, tool call PASS, vision PASS.
Reasoning that stops: all nine deterministic tasks pass at default, low, medium, and xhigh.
Multimodal preserved: this remains an image-text-to-text checkpoint, not a text-only conversion.
Reproducible: the exact parent revision, modified tensor list, and evidence hashes ship with the model.

Bounded reasoning

The upstream template defaulted unspecified reasoning to xhigh. This release defaults to medium and adds a stop condition to xhigh: once an answer is established, the model must stop instead of repeating or restarting its deliberation. The checkpoint was tested across 36 deterministic effort/task combinations; all 36 terminated correctly.

Explicit controls remain available through chat_template_kwargs:

python
extra_body={"chat_template_kwargs": {"enable_thinking": True, "reasoning_effort": "xhigh"}}

Behavior and capability

Table with columns: Gate, Result
Gate	Result
Parent refusal	30 / 32 (93.75%)
SuperQwen refusal	0 / 32
Empty output	0 / 32
Capability	7 / 8 (paired-parent floor)
Tool use	PASS
Vision	PASS
Overthinking	36 / 36 PASS

Precision and integrity

Full BF16 checkpoint; this repository is not quantized.
100 declared tensors changed and zero unexpected tensors changed.
Vision (333 tensors) and MTP (15 tensors) remain byte/value exact.
Parent revision is pinned to 1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0.

Decode performance

Measured on one DGX Spark with fixed-length generation and the sparkDash-style post-first-token contract:

Table with columns: Prompt / concurrency, Aggregate decode
Prompt / concurrency	Aggregate decode
p256 / C1	4.3411 tok/s

Verified long context

The official native limit is 262,144 tokens. A 262,043-token prompt completed end to end and retrieved its hidden needle. Long-context acceptance is not a claim of perfect recall on every task.

Serving

bash
QWEN38_SPECULATIVE_TOKENS=0   bash repro/scripts/serve_superqwen38_replica.sh   /model SuperQwen3.8-27b-abliterated 8888

Uncensored behavior

“Abliterated” means that the measured refusal direction was reduced. It does not mean that every response is correct, harmless, or suitable for every deployment. Operators remain responsible for access controls and downstream safeguards.

Limitations

Abliteration changes refusal behavior and may surface content the parent declined.
The capability, tool, vision, and overthinking suites are finite regression gates.
Speed is hardware- and runtime-specific.
The 1M expansion run was explicitly stopped before completion and is not claimed as a pass.

Evidence identities

Table with columns: Evidence, SHA-256
Evidence	SHA-256
abliteration verification	`6d7b8b9abc431ed3845b18438fc48a6cc169e2ba93f83ada54174ae93765a61e`
abliteration recipe	`3c057cf9364e72857d9a48b6c8888c6b438cbfb519459dcdfbf23b22c2bc6796`
842-pair OBLITERATUS corpus manifest	`be770a8469fafe7a9a76db5d82a7893c706c4414e1de6089f1d329b906480aa8`
bounded-thinking template	`352baa520d23c77129d1cb9b007c5480118c174e12e271457ffe915fb466f3fb`
parent refusal baseline	`7a4177613c86ae0db589cb4fc167ff8a0d45b0d49f44e40b90c9405f4b2b9bb6`

License

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

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Model Details

Model Provider

vishrutJ

Model Tree

Base

Qwen/Qwen3.8-27B

Fine-tuned

this model

Input Modalities

TextImageVideo

Output Modalities

Text

Supported Functionality