FlashAlign 1.1.1

Adds TensorFold 1.2.0 support for the native DeepSeek starter.

• Uses the same profile payload as the native vLLM backend.
• Keeps speculative decoding active and stock/profile caches separate in RAM.
• Adds profile verification and managed setup for the TensorFold runtime.

The TensorFold path requires the pinned LibertAIDAI EXL3 3.5-bpw checkpoint, Linux x86_64, four SM120 GPUs and strength 1.0. JSON-schema and guided output are unsupported.

Setup and requirements · Validation results

FlashAlign 1.1.0

Replaces the DeepSeek starter with the native Keys profile.

• Applies the source’s 26 WO-B matrices and 26 scale tensors per authorized request.
• Adds the pinned native vLLM backend for the official DeepSeek V4.1 Flash 1017 FP8 checkpoint.
• Retrieves the 1.09 GB profile payload from this public release with size and SHA-256 verification.
• Preserves speculative decoding and separate normal/profile caches.

Version 1.1.0 supports this native profile on vLLM. Version 1.1.1 adds TensorFold support. SGLang and llama.cpp do not implement this operator. Existing compact-profile integrations remain available for compatible profiles.

Source edit: drowzeys/keys-DeepSeek-V4.1-Flash-1017-Abliterated. Runtime packaging by Aevonix. The source MIT notice and responsible-use agreement apply to the profile; FlashAlign software remains AGPL-3.0-only with a commercial license option.

Profile requirements · Validation results and limits

FlashAlign 1.0.0

ReAlign models instantly, on demand.

FlashAlign applies alignment profiles per request to supported base and quantized models. Use it directly, with Hermes or OpenClaw, or through a custom harness. Choose verified contacts or Always On, and control tool and subagent inheritance.

This release includes experimental starter profiles for DeepSeek V4.1 Flash 1017, GLM 5.3 Full, GLM 5.3 Flash, Qwen3.8-27B and Gemma 4 26B A4B. Supported engines vary by model; see the compatibility table before downloading model weights.

• Get started
• Model downloads
• Measured results and limitations
• Demo notes and prompt attribution
• Starter-profile sources and terms

Model weights are downloaded separately. The source archive includes all five starter profiles. The wheel provides the Python client; TensorFold runtime downloads are selected by the setup wizard for the supported host and model.

The starter profiles do not guarantee zero refusals or exact reproduction of an edited checkpoint. The included profiles focus on refusal reduction; refusal-increasing profiles are not included.

FlashAlign software is open source under AGPL-3.0-only. Commercial licenses are available from Aevonix Inc. Model and third-party components retain their own license terms.

Verify downloaded files against SHA256SUMS. RELEASE-MANIFEST.json records the release commit, runtime provenance and profile hashes.