--- license: other license_name: swift-open-license-1.0 license_link: LICENSE library_name: transformers pipeline_tag: image-text-to-text base_model: - d0xin/Swift-1.5-Qwen3.8-Flash-Next-NVFP4-FP8PLE tags: - qwen - qwen3.8 - flash-next - swift - nvfp4 - fp8 - fp8-ple - abliteration - uncensored - safetensors - sglang - pennyroyal - blackwell - local-llm --- > Related models: [all models](https://huggingface.co/d0xin) > > [Swift-1.5-Qwen3.8-27B-Uncensored-BF16](https://huggingface.co/d0xin/Swift-1.5-Qwen3.8-27B-Uncensored-BF16) · > [Swift-1.5-Qwen3.8-27B-Uncensored-FP8](https://huggingface.co/d0xin/Swift-1.5-Qwen3.8-27B-Uncensored-FP8) · > [Swift-1.5-Qwen3.8-27B-Uncensored-FP8-NInfer](https://huggingface.co/d0xin/Swift-1.5-Qwen3.8-27B-Uncensored-FP8-NInfer) · > [Swift-1.5-Qwen3.8-Flash-Next-NVFP4-FP8PLE](https://huggingface.co/d0xin/Swift-1.5-Qwen3.8-Flash-Next-NVFP4-FP8PLE) · > [Rank-2 Abliteration Patch](https://huggingface.co/d0xin/Swift-1.5-Qwen3.8-Flash-Next-Rank2-Abliteration-Patch) # Swift-1.5 Qwen3.8 Flash-Next NVFP4 FP8PLE — Rank-2 Abliterated Full baked checkpoint of the validated Rank-2 refusal-direction ablation for: `d0xin/Swift-1.5-Qwen3.8-Flash-Next-NVFP4-FP8PLE` This release contains the **complete ready-to-serve model weights**. No offline patching or runtime abliteration hook is required. The checkpoint preserves the FP8 PLE conversion and NVFP4 layout of the parent release while baking the validated Rank-2 `{Orca, Swift⊥}` projection directly into the affected model weights. > **Important:** do not apply the separate Rank-2 patch or the runtime `flashnext_abliteration` hook on top of this checkpoint. The transformation is already baked into the weights. > **Hardware / deployment note:** this full baked checkpoint is approximately **126 GiB** and is **not intended to be loaded as a complete in-memory checkpoint on a single NVIDIA RTX PRO 6000 96 GB**. > > For a single RTX PRO 6000 96 GB, use the validated deployment path based on the parent [`Swift-1.5-Qwen3.8-Flash-Next-NVFP4-FP8PLE`](https://huggingface.co/d0xin/Swift-1.5-Qwen3.8-Flash-Next-NVFP4-FP8PLE) together with the published Rank-2 runtime patch and NVMe-PLE setup. > > This full baked release is intended primarily for **multi-GPU or larger-memory deployments, runtimes with explicit model sharding/offload support, further conversion or quantization, and reproducible distribution of the already-abliterated weights**. ## Model lineage ```text Qwen/Qwen3.8-Flash-Next ↓ ukisai/Swift1.5-Qwen3.8-Flash-Next ↓ ukisai/Swift-1.5-Qwen3.8-Flash-Next-NVFP4 ↓ d0xin/Swift-1.5-Qwen3.8-Flash-Next-NVFP4-FP8PLE ↓ Rank-2 {Orca, Swift⊥} refusal-direction ablation ↓ d0xin/Swift-1.5-Qwen3.8-Flash-Next-NVFP4-FP8PLE-Rank2-Abliterated ``` ## What changed The transformation suppresses the learned refusal direction while keeping the parent checkpoint architecture and quantization layout intact. | Property | Value | |---|---:| | Projection basis | `{Orca, Swift⊥}` | | Rank | **2** | | Alpha | **1.0** | | BF16 targets | **100** | | NVFP4 expert weights transformed | **24,576** | | Companion scale tensors | **49,152** | | Modified safetensors shards | **49** | | Modified payload | **78.227 GiB** | | FP8 PLE shards | **10 — unchanged** | The large PLE embedding table remains in the FP8 E4M3 representation from the parent FP8PLE release. The abliteration does not rebuild or modify those 10 PLE shards. ## Artifact This repository contains the full baked checkpoint. | Property | Value | |---|---:| | Repository files | **74** | | Safetensors shards | **59** | | Indexed tensors | **296,475** | | Published payload | **135,257,737,044 bytes** | | Safetensors payload | **135,195,697,746 bytes** | | Configured context | **262,144 tokens** | | PLE | **FP8 E4M3** | | Target weights | **NVFP4 / mixed precision** | The Hugging Face upload was verified after transfer: - **74 / 74** repository files present - **59 / 59** safetensors shards present - remote safetensors byte count exactly matches the local checkpoint - required config, tokenizer and safetensors-index files present ## Abliteration validation The baked checkpoint was independently compared with the validated reference transformation. Reference-build validation: - 59 / 59 safetensors shards - 296,475 / 296,475 indexed tensors - 0 missing tensors - 0 extra tensors - 0 duplicate tensors - 0 wrong-shard mappings - 49 / 49 modified-shard SHA256 checks passed Tensor-level differential audit: ```text expected_touched = 1538 touched_changed = 1538 touched_same = 0 unexpected_changed = 0 ``` All expected tensors changed and no unrelated tensor changed. This establishes that the offline baked checkpoint reproduces the validated Rank-2 transformation rather than being an independently tuned approximation. ## Refusal behavior The validated Rank-2 transform produced: | Evaluation | Result | |---|---:| | Primary validation set | **80 / 80 DIRECT — 100%** | | Held-out JBB non-AdvBench set | **79 / 80 DIRECT — 98.75%** | | Held-out refusals | **1 / 80 — 1.25%** | The held-out evaluation used 80 prompts kept separate from the prompts used while developing the projection. Configuration: - temperature: `0` - maximum output: `192` tokens - concurrency: `4` - deterministic refusal-marker classifier - output classified as `REFUSE` when a refusal marker was detected; otherwise `DIRECT` These numbers describe this fixed evaluation only. They are not a guarantee that the model will never refuse under another prompt, system message, sampling configuration or runtime. ## Performance Reference generation-speed runs with the Rank-2 transform on a single NVIDIA RTX PRO 6000 Blackwell 96 GB: | Run | Throughput | |---|---:| | 1 | **141.75 tok/s** | | 2 | **138.66 tok/s** | | 3 | **154.47 tok/s** | | Median | **141.75 tok/s** | Runtime: - Pennyroyal / SGLang v2.5.3 - single NVIDIA RTX PRO 6000 Blackwell 96 GB - native NEXTN / MTP speculative decoding - FP8 KV cache - FP8 PLE - NVMe-backed PLE - 262,144-token configured context Performance is hardware-, runtime- and workload-dependent. ## Functional regression checks The validated Rank-2 deployment passed: - normal chat-completion inference - native NEXTN / MTP speculative decoding - structured startup warmup - tool calling with correctly formed function arguments - end-to-end routing through `hybrid_auto` - 262,144-token configured context - NVMe-backed FP8 PLE operation No agent/tool-calling regression was observed in these smoke tests. ## Validation provenance The behavioral, performance and functional measurements above were originally performed with the validated runtime application of the same Rank-2 `{Orca, Swift⊥}` projection. The offline baker was then independently checked against the baked reference checkpoint: - all 49 modified shard SHA256 hashes matched - transformation totals matched exactly - all 1,538 expected tensors changed - zero unrelated tensors changed This repository contains that baked transformation. The behavioral benchmark was not separately rerun merely as a consequence of uploading the baked checkpoint to Hugging Face. ### Parent-model quality reference The parent [`d0xin/Swift-1.5-Qwen3.8-Flash-Next-NVFP4-FP8PLE`](https://huggingface.co/d0xin/Swift-1.5-Qwen3.8-Flash-Next-NVFP4-FP8PLE) qualification produced: | Test | Result | |---|---:| | MMLU-Pro | **226 / 280 — 80.71%** | | Native configured context | **262,144 — PASS** | | Retrieval near context limit | **~261.7K input tokens — PASS** | | Agentic tool/workflow smoke | **7 / 7 PASS** | These are **parent-checkpoint reference results**, not a post-abliteration MMLU-Pro score. A separate post-abliteration capability suite can be reported when measured directly on this baked release. ## Deployment notes The original Rank-2 validation used Pennyroyal / SGLang v2.5.3 on NVIDIA Blackwell. The documented single-RTX-PRO-6000 results were obtained with the parent FP8PLE checkpoint plus the Rank-2 runtime transformation and NVMe-backed PLE path. They should **not** be interpreted as evidence that this approximately 126 GiB baked checkpoint can be loaded entirely into the 96 GB VRAM of one RTX PRO 6000. Reference image: ```text ghcr.io/jpezzulli/sglang-rtxpro6000:v2.5.3 ``` Download the checkpoint: ```bash hf download \ d0xin/Swift-1.5-Qwen3.8-Flash-Next-NVFP4-FP8PLE-Rank2-Abliterated \ --local-dir /srv/models/Swift-1.5-Qwen3.8-Flash-Next-NVFP4-FP8PLE-Rank2-Abliterated ``` Use the Pennyroyal `next-plain` profile and point `TARGET_MODEL` directly at the baked checkpoint: ```text PENNYROYAL_PROFILE=next-plain PENNYROYAL_IMAGE=ghcr.io/jpezzulli/sglang-rtxpro6000:v2.5.3 TARGET_MODEL=/models/Swift-1.5-Qwen3.8-Flash-Next-NVFP4-FP8PLE-Rank2-Abliterated PENNY_PLE_BACKEND=nvme PENNY_PLE_NVME_MODEL=/models/Swift-1.5-Qwen3.8-Flash-Next-NVFP4-FP8PLE-Rank2-Abliterated-nvme SGLANG_SM120_ONLINE_MXFP8=true SGLANG_MM_PREPROCESS_DEVICE=cpu MAX_RUNNING_REQUESTS=4 MAX_MAMBA_CACHE_SIZE=24 ``` ### Prepare the NVMe PLE overlay The NVMe PLE derivative is a machine-local runtime artifact and is not included in this repository. Prepare it once from the baked checkpoint: ```bash docker compose run --rm --no-deps \ -v /srv/models:/models \ pennyroyal exec .venv/bin/python \ scripts/pennyroyal/prepare_ple_nvme.py \ --source /models/Swift-1.5-Qwen3.8-Flash-Next-NVFP4-FP8PLE-Rank2-Abliterated \ --output /models/Swift-1.5-Qwen3.8-Flash-Next-NVFP4-FP8PLE-Rank2-Abliterated-nvme ``` Then start Pennyroyal normally. ### Do not double-apply the abliteration This is a **baked** model. Do not enable: - the Rank-2 patch baker - `flashnext_abliteration.pth` - another runtime `{Orca, Swift⊥}` projection hook when serving this checkpoint. Those mechanisms are intended for applying the transformation to the unabliterated FP8PLE parent. Applying them again would transform already modified weights a second time and is not the model validated here. ## Reproducibility The complete reproducible transformation kit is published separately: [`d0xin/Swift-1.5-Qwen3.8-Flash-Next-Rank2-Abliteration-Patch`](https://huggingface.co/d0xin/Swift-1.5-Qwen3.8-Flash-Next-Rank2-Abliteration-Patch) It contains: - `direction_rank2_orca_swift.pt` - `plan.json` - `flashnext_abliteration_stage_b_rank2.py` - `bake_flashnext_rank2.py` - `apply_patch.sh` - `reference_BAKE_MANIFEST.jsonl` - `BASE_FINGERPRINTS.json` - `SHA256SUMS` Reference hashes: | Artifact | SHA-256 | |---|---| | Direction | `8eb11e23dcbea3cbc040f5077856b19ea475d5af81d1a62f789ebe96d5130b2c` | | Plan | `22ab558cf8708bf1f6a57d475ce2bc0c9ce7b5a9a54e5f756be70e28f30a7f50` | | Stage-B implementation | `5138983ee95bd80b2645ce8f37ff20bab0c023376f6f68ee0c567842a6ebfbcf` | | Baker | `e777252cdb7fa445c4d19cd316f3218d736de9af2a31e9b4ba56b9c6322f8e8c` | ## Related releases - [`Swift-1.5-Qwen3.8-Flash-Next-NVFP4-FP8PLE`](https://huggingface.co/d0xin/Swift-1.5-Qwen3.8-Flash-Next-NVFP4-FP8PLE) — unabliterated FP8-PLE parent - [`Swift-1.5-Qwen3.8-Flash-Next-Rank2-Abliteration-Patch`](https://huggingface.co/d0xin/Swift-1.5-Qwen3.8-Flash-Next-Rank2-Abliteration-Patch) — reproducible Rank-2 patch kit - [`Swift-1.5-Qwen3.8-Flash-Next-NVFP4-FP8PLE-Rank2-Abliterated`](https://huggingface.co/d0xin/Swift-1.5-Qwen3.8-Flash-Next-NVFP4-FP8PLE-Rank2-Abliterated) — this full baked checkpoint ## Safety and responsible use This model intentionally has substantially reduced refusal behavior. Abliteration changes model-level refusal behavior; it does not provide application-level safety controls and it does not remove knowledge from the model. Deployment operators are responsible for deciding what moderation, access control, monitoring and other safeguards are appropriate for their use case. The refusal measurements above describe fixed evaluations and should not be interpreted as a universal guarantee about model behavior. ## License and provenance This model derives from Swift 1.5 Qwen3.8 Flash-Next. UkisAI's Swift contribution is distributed under the Swift Open License v1.0. The underlying Qwen3.8-Flash-Next portions remain subject to the Qwen Community License 1.0. See the included `LICENSE`, `LICENSE-QWEN` and `NOTICE` files for governing terms and attribution requirements. The Rank-2 transformation in this repository modifies the parent checkpoint; it does not replace or supersede upstream license terms.