ReAlign models instantly, on demand.

By Aevonix Research · Version 1.2.0

▶ Watch the 40-second video with sound · About this video

Get started · Setup · Compatibility · Docs

Apply alignment profiles per request. Use abliteration for approved users and channels, or keep a profile active across every routed request.

FlashAlign works with supported base and quantized weights without permanently changing them. Run it directly, with Hermes or OpenClaw, or through your own harness. Managed inference runs on Linux with NVIDIA GPUs. See compatibility for DGX Spark support and engine requirements.

The starter profiles focus on refusal reduction. Future profiles could also increase refusals for public or higher-risk tasks.

Restricting abliteration to verified users and channels reduces the risk of an abliterated model being abused through malicious prompt injection. Tool outputs within an authorized request still rely on the harness's protections.

Get started

Download a supported model first. The model download guide has pinned commands for all five starter profiles and their supported formats. The wizard detects local models and harnesses, selects a matching profile, prepares the engine and creates a deployment signing key.

git clone https://github.com/Aevonix/FlashAlign.git
cd FlashAlign
bash ./install.sh --auto

Choose Verified contacts or Always On. In Verified contacts mode, choose whether tools inherit the profile. Setup checks activation and backs up changed configuration. Chat integrations restart their selected gateways; local terminal setup prints its launch command.

Five starter profiles are available. Four compact profiles are included in the checkout. The DeepSeek profile uses a 1.09 GB native tensor payload; setup fetches it from a public release and verifies its checksum. No private repository or login is required. Profile details.

Use the profile inspector to check a local profile's model pin, operators, sites, ranks and hashes before setup.

Experimental test profiles based on Keys, OrcaRouter, Dealign, and MoonMath abliterated checkpoints. Fitting and runtime packaging by Aevonix. See the exact checkpoint versions.

Model	Supported engines
DeepSeek V4.1 Flash 1017	vLLM, TensorFold
GLM 5.3 Full	TensorFold
GLM 5.3 Flash	TensorFold, SGLang
Qwen3.8-27B	TensorFold, vLLM, SGLang, llama.cpp
Gemma 4 26B A4B	TensorFold, vLLM, SGLang, llama.cpp

See compatibility for hardware requirements, pinned checkpoints and feature limits, and validation for tested configurations and results.

Switching happens per request. Engines that advertise profile hot loading can also add compatible profiles without a restart. Loading another model still requires its own engine setup. See profile management.

Use --model-path /path/to/model to select a model. Run bash ./install.sh to connect to an engine that is already running. See the engine selection guide for supported model and engine combinations.

For standalone setup with a local Qwen checkpoint:

bash ./install.sh --engine-only --backend vllm --model-path /models/qwen38-27b --yes

See direct access for the Python client and custom harnesses for integration hooks.

To use profiles from another folder:

bash ./install.sh --profiles /path/to/profiles

To inspect your setup first:

bash ./install.sh --detect

The setup guide covers connection and activation. The engine's effective-profile response confirms that the selected profile is active.

Check and manage your deployment

After setup, run:

bash ./flashalign.sh status
bash ./flashalign.sh doctor --probe
bash ./flashalign.sh dashboard
bash ./flashalign.sh profile list

Status reads runtime metadata. Doctor sends a synthetic request only with --probe. The dashboard shows measured values when the engine provides them. See operations, profile management and benchmarks.

Measured results

On Qwen3.8-27B, the automated FR443 judge counted 289 refusals in 443 stock responses and 11 in 443 profile responses. The suite includes safe, harmful and unsafe requests. Human audit and broader capability testing remain pending. Full results and settings.

A separate Qwen/SGLang timing study measured 1.324× end-to-end completion-token throughput for the opt-in graph Owner Mode path compared with the previous release's eager path. This was a fixed 256-token workload with three pairs of engine starts, not a decode-speed measurement. Timing results and reproducible data.

A fixed follow-up passed completion and GSM8K, but did not establish all capability gates. Eager remains the default and graphs remain opt-in. Quality results and limits.

Choose when to apply FlashAlign

Verified contacts is the default. It applies a profile to messages from an approved channel, account and contact.

Always On applies the configured profile to every request routed through FlashAlign, including tool and subagent requests. It overrides per-tool alignment settings and stays selected across restarts. The harness still decides who can send messages and which tools can run. Always On does not mark callers as verified. Requests fail if the configured profile cannot be applied.

To set up Always On, run bash ./install.sh --auto --mode always_on. Contact approval is not required in this mode.

Point custom harnesses and OpenAI-compatible clients at the Always On gateway. FlashAlign controls only the model calls sent through its routes or gateway URL.

Switch an existing harness deployment without a restart:

python3 hermes-owner-mode/manage.py --store ~/.config/owner-mode/approvals.json \
  mode always_on

Use mode verified to return to contact-based access. The selection survives restarts and does not re-enable a disabled deployment. For an engine-only deployment, change its mode through the wizard's activation step; see direct access.

Built-in messaging adapters cover WhatsApp and RCS direct messages, plus opt-in SMS and Telegram. On supported messaging transports, photos, voice notes, attachments, forwarded or quoted content, and queued message groups retain their sender verification through the harness.

Explicitly authorized local terminal sessions work with Hermes and OpenClaw and are text-only in V1. Other transports can use the adapter interface.

python3 hermes-owner-mode/manage.py --store ~/.config/owner-mode/approvals.json \
  approve whatsapp +15550000011 --account default

Replace the example number and account with the IDs used by your harness. Approvals survive restarts and remain active until revoked. Each new turn must arrive through the verified channel. A sender ID written in a message grants no access.

To revoke contact approval:

python3 hermes-owner-mode/manage.py --store ~/.config/owner-mode/approvals.json \
  revoke whatsapp +15550000011 --account default

In Verified contacts mode, revocation applies to subsequent model requests, including delegated work. It does not turn off Always On.

Choose which tools inherit the profile

In Verified contacts mode, choose the default during setup and override individual tools by name:

python3 hermes-owner-mode/manage.py --store ~/.config/owner-mode/approvals.json \
  tool delegate_task off
python3 hermes-owner-mode/manage.py --store ~/.config/owner-mode/approvals.json \
  tool sessions_spawn on

Enabled tools can pass the active profile to verified child work. Disabled tools use the standard model and pass no profile to their children. Always On applies the profile to every routed tool and subagent request, regardless of these flags. The settings are saved for a later switch back to Verified contacts. Your harness controls tool execution permissions in both modes.

Compatibility

Component	Requirement
Inference host	Linux and NVIDIA GPU; matched CPU and GPU architecture for TensorFold bundles
Python	3.10 or later for setup and adapters
Hermes	Native ingress, request and tool-execution hooks
OpenClaw	Compatible native plugin API; Node 24.22.1+ on 24.x or 26.12.1+ on 26.x
Model API	OpenAI-compatible Chat Completions and the FlashAlign engine contract
Model profile	A matching profile loaded by the engine

See compatibility for channel, engine and delegation requirements.

Documentation

• Setup and activation
• Status, doctor and dashboard
• Live profile management
• Managed engine setup
• Agent-assisted setup
• Harness and model compatibility
• Validation and tested configurations
• Measured results
• Run your own benchmarks
• Native message and media bindings
• Channel adapters and terminal access
• Hermes configuration
• OpenClaw adapter
• Engine API
• Extension interfaces
• Security model and vulnerability reports
• Development and tests
• Vector logos and brand kit

License

FlashAlign software is open source under AGPL-3.0-only. Commercial licenses are available from Aevonix Inc for closed-source products and hosted services. AGPL also permits commercial use when its conditions are met. Model, profile and dependency terms remain separate; see third-party notices.

Contributions are accepted under the contribution terms, which grant Aevonix commercial and sublicensing rights.

Built by Aevonix Research.