MODEL DISCLOSURE
Qwen3.8-27B Abliterated.
Exact artifact, AWQ W4A16 quantization, RTX 5090 hardware, context, and capability limits of the AblitAI text endpoint.
Private beta. Bitcoin checkout open (beta); card closed. No uptime SLA and no published performance benchmarks.
Current serving envelope
| Stable model ID | qwen3.8-27b-abliterated |
|---|---|
| Community artifact | twolven/Qwen3.8-27B-abliterated-AWQ-MTP |
| Pinned revision | c2f5341e7c9a1c55e6d61cfc8e6d0ca897bd7443 |
| Quantization | AWQ W4A16, group size 128 · bf16 compute · fp8 KV cache |
| Serving engine | vLLM 0.29.0 |
| Launch GPU | NVIDIA RTX 5090, 32GB-class VRAM |
| Total context | 131,072 tokens, including prompt, template, and completion |
| Maximum completion | Up to 131,072 tokens (default 512 if max_tokens omitted) |
| Simultaneous sequences | Two |
| Modality / default | Text / non-thinking (vision is a separate routed model) |
| Inactive features | MTP / speculative decoding; qwen3_coder parser (tools are otherwise on) |
Origin and independence
This is a community-modified artifact from the Qwen model family, independently served by AblitAI. No affiliation or endorsement by Alibaba or Qwen is implied. Apache-2.0 attribution and applicable artifact notices remain relevant; a model license does not itself grant general trademark rights.
What works here
Text Chat Completions and SSE streaming are implemented. The exact GPU passed bounded text and streaming checks followed by more than 24 continuous hours without an unexplained restart or memory failure. These are operational observations, not an uptime promise or published benchmark.
Tools are on. This card is the text endpoint. Multimodal image chat is available on the API via a separate vision model; see the vision model card. Audio and thinking requests remain unsupported on this text endpoint. Coding harnesses stay reference-only; the support matrix separates local client evidence from sustained production claims. The artifact's MTP suffix does not mean speculative decoding is active.
What quantization changes
The selected community artifact describes llm-compressor 0.13.0, asymmetric W4A16 weights and group size 128, calibrated on 128 UltraChat samples of 1,024 tokens. Lower-precision weights reduce memory use and can change output quality. They do not determine the usable context window by themselves. Read the deployment explanation.
Behavior and lawful use
“Abliterated” refers to a community modification intended to reduce some refusal tendencies. It guarantees neither correctness nor compliance with every instruction. The Acceptable Use Policy still applies; model modifications do not waive lawful-use restrictions.
Change policy
The serving process is pinned to the disclosed revision rather than tracking a moving branch. Model, quantization, context, parser, or engine changes require fresh acceptance checks. Future capacity is discussed in the roadmap without promised dates. Check live status and the changelog for current service context.
Next: make a text request or understand token accounting.
Transient prefix state
The current model runtime can reuse transient in-memory prefix state across requests. This state is distinct from persistent prompt or response logs and databases. Disabling prefix caching is prepared but has not yet been deployed; content should not be assumed to clear from GPU memory immediately after a request.
Runtime telemetry opt-outs are prepared but not yet applied. Request, output, statistics, and HTTP access logging are disabled. Operational telemetry is a separate boundary from prompt and response logs; no claim is made here that every runtime telemetry mechanism is disabled.