Qwen3.8-27B-GSQ3-NInfer

모델 소개

The repository contains a verbatim repack of the 3-bit ISTA-DASLab/Qwen3.8-27B-3Bit-GSQ checkpoint into the native NInfer `.ninfer` artifact format, sized to run the full 100,000-token, vision, and speculative-decoding profile on a single 16 GB **RTX 4080**. It is intended only for the RTX 4080 NInfer fork; it is not a Transformers checkpoint, Safetensors distribution, or GGUF file. The Text body is 320 matrices in the registered `Q3G128_F16S` scheme: symmetric 3-bit codes in `[-4, 3]`, one FP16 multiplier per 128-value group, 3.125 bits per weight. The token embedding, full output head, and optimized draft head use `Q4G64_F16S`; the MTP layer and Vision tower follow NInfer's registered `groupwise-int` recipes; the DFlash2 companion was requantized to `Q4G64_F16S`. The file also carries the tokenizer, chat-template, generation, and media-processor objects required by NInfer. The 323 packed matrices that come from the publisher's checkpoint — the 320 Text-body matrices, the token embedding, the full output head, and the draft head (a row gather of the output head) — are a lossless repack: every code and every group scale was copied unchanged, with only the publisher's shifted bit-plane transform inverted (`compressed-tensors` stores the unsigned `code + 2^(bits-1)` form; the artifact stores two's-complement fields). No code was requantized. The only value deviation from the source is **218 of 240,271,360** multipliers: bf16 subnormal words whose FP16 rounding error is at most `2**-25` (max observed `2.98e-8`), audited per group in the conversion report. The publisher's task evaluations therefore describe the represented Text and vocabulary weights.

주요 특징

特点摘要待依据来源补齐，不凭模型名称推断能力。

한계 및 유의 사항

구체적 한계를 판단할 출처가 부족합니다. 실제 성능, 지연, 데이터 정책, 라이선스와 버전 수명을 확인하세요.

출처별 요금 및 비용 분석

요금은 채널, 지역, 설정에 따라 다릅니다. 각 채널 요금은 개별 표시되며 제공사의 통일 공식 요금이 아닙니다.

공개 평가 및 실행 설정

暂无已采集的可核验公开评分。模型未上榜不代表能力较低。

자주 묻는 질문

출처 및 추적 기록

许可证：apache-2.0。模型介绍与第三方评测均以原来源的具体版本和方法为准。