Heretic is live · free · last verified 8 Nov 2026 · viability monitor
Heretic
Open-source tool that removes refusal behaviour from open-weight language models so they follow your instructions
A genuinely useful, honestly scoped tool: one job, free, AGPLv3, and the docs actually cover installation and security. Pick it when you already host open-weight models and want refusal behaviour gone today instead of after a training run. Pass if you need a vendor to indemnify the output or if AGPL source-disclosure obligations would be a problem for your deployment.
Verdict generatedVerified 5d ago · liveness 69/100 · cite: rightaichoice.com/tools/heretic
- AI safety researchers studying how refusal and censorship training shapes model behaviour
- Developers running local open-weight inference who need unfiltered output from weights they already host
- Technical users comfortable with pip, Python environments and command-line model tooling
- Open-source contributors and hobbyists who want a maintained alternative to writing abliteration scripts
- Non-technical users who want a packaged app with a GUI and no Python setup
- Organisations that need vendor liability cover, indemnification or a support SLA
- Teams in regulated industries requiring documented compliance sign-off on model behaviour
We scan live Reddit threads, YouTube comments, X posts, G2 reviews and other communities — and hand you an honest verdict in under a minute.
- Honest verdict, not marketing
- Real pros & cons from real users
- Attributed quotes with receipts
3 free scans · no card needed
Skip Heretic if you need a supported commercial product with an SLA, or if you cannot modify and self-host the open-weight model you want to use.
The software is free, but the GPU time and hosting to run an open-weight model locally are your own cost.
Heretic costs nothing to use — it is free software under the AGPLv3-or-later licence. The real spend comparison is against your compute: it sits alongside free self-hosting stacks, and well below the per-seat cost of hosted commercial LLM platforms. Budget for a GPU you control, not a subscription.
In short
Heretic — Open-source tool that removes refusal behaviour from open-weight language models so they follow your instructions. Best for AI safety researchers studying how refusal and censorship training shapes model behaviour, Developers running local open-weight inference who need unfiltered output from weights they already host, Technical users comfortable with pip, Python environments and command-line model tooling. Free to use.
What is Heretic?
Heretic — Open-source tool that removes refusal behaviour from open-weight language models so they follow your instructions. Best for AI safety researchers studying how refusal and censorship training shapes model behaviour, Developers running local open-weight inference who need unfiltered output from weights they already host, Technical users comfortable with pip, Python environments and command-line model tooling. Free to use.
Heretic is a free, AGPLv3-or-later command-line utility that strips refusal behaviour out of open-weight language models. You install it with a single pip command, point it at a model, and it goes to work modifying the weights so the model answers the prompt you actually gave instead of deflecting it. The project is maintained by Philipp Emanuel Weidmann and contributors, and the live site names Qwen/Qwen3.5-4B as its worked example. The intended audience is narrow on purpose: AI-safety researchers who want to study how refusal training shapes model behaviour, developers running local inference who need unfiltered output from weights they already host, and people comfortable enough with a Python environment to debug a dependency. The project documents Tutorial, Installation, Configuration and Security pages, and exposes both a CLI entry point (`heretic Qwen/Qwen3.5-4B`) and a Python API for scripted or pipeline use. Mechanically, Heretic patches or wraps an existing model rather than replacing it, with the stated aim of preserving performance and coherence. It runs locally on your own hardware, collects no telemetry, and ships its source on GitHub with model artefacts on Hugging Face. Support runs through Discord and Matrix rather than a ticketing desk. The licence is the part buyers skip and later regret: AGPLv3-or-later means anyone serving a modified model over a network inherits source-availability obligations. Compare it to bespoke fine-tuning or abliteration scripts off GitHub — Heretic is the packaged, maintained version of that work, and it is the free option against paid model-modification or hosted-unfiltered-endpoint services.
Key features
- Remove refusal and censorship behaviour from open-weight language models
- Patch or wrap an existing model rather than replacing it
- Preserve model performance and coherence during modification
- Command-line interface: heretic Qwen/Qwen3.5-4B
- Python API for scripted and pipeline use
- Install via pip install -U heretic-llm
- Documented Tutorial, Installation, Configuration and Security pages
- Runs locally on your own machine
- No telemetry or data collection
- AGPLv3-or-later free software licence
- Source code on GitHub, model artefacts on Hugging Face
- Community support via Discord and Matrix
- Works with open-weight models such as Qwen/Qwen3.5-4B
Behind the verdict
We'd reach for Heretic when the bottleneck is time, not skill. If you already know what an open-weight model is and you have a Python environment that works, the install is two lines and the configuration docs are real. That is a shorter path than writing your own abliteration script and debugging weight surgery alone. Where it bites is scope. Heretic modifies weights you supply; it does not host anything for you, and it does not give you an endpoint. Everything downstream — inference server, guardrails, logging, evals — is your problem. Teams expecting a drop-in API will be disappointed within an hour. The AGPLv3-or-later licence deserves more attention than it usually gets. Running it locally for research is one thing. Wrapping a modified Heretic model in a network service is another, and the source-availability obligation follows you. Legal should read that before engineering ships it. Against the closest alternative — hand-rolled fine-tuning or a borrowed abliteration notebook — Heretic wins on maintenance and loses on flexibility. Fine-tuning lets you target specific behaviours; Heretic gives you the blunt instrument and the docs to swing it. For researchers comparing censorship-training regimes across many models, blunt and repeatable is the right trade. In practice the output quality question is the one we'd test first. The project claims to preserve performance and coherence, and that is worth verifying on your own evals before you commit a product to it. Refusal removal can shift behaviour in ways unrelated to refusals. Support is Discord and Matrix. That is fine for hobbyists and academic labs; it is not a contract. If something breaks the week before a deadline, you are reading source. Who should skip it outright: non-technical users wanting a packaged
Models under the hood
as of 2026-11-13
Limitations
- Heretic is a command-line tool distributed as AGPLv3-or-later free software; you install it with 'pip install -U heretic-llm' and need to be comfortable with Python tooling and terminal workflows.
- The live homepage names only one example model (Qwen/Qwen3.5-4B), so the full range of supported open-weight models is not stated in the evidence.
- The project relies on community channels (Discord, Matrix) rather than commercial support, and no SLA or liability arrangement is documented.
- Because the licence is AGPLv3-or-later, running a modified Heretic as a network service carries a source-availability obligation you should have reviewed by counsel.
as of 2026-10-08
Researching Heretic? Get your full AI stack in 60 seconds.
Free, no signup — tell us your goal and get tools matched to your budget & existing stack.
Is Heretic still active?
Heretic is live · free · last verified 8 Nov 2026 · viability monitor.
Viability score
How well maintained and how widely used is Heretic? Built from what the vendor actually publishes (docs, changelog, tutorials, integrations, pricing), whether the site is live, and how much real users discuss it. How we calculate this
Last calculated: November 2026
How we score →What do users say about Heretic?
What people actually say about Heretic — is it worth it?
We scanned public community sources for Heretic on Nov 6, 2026 and could not establish that the discussion we found is about this tool rather than something else sharing its name. Only 2 of the posts we fetched could be positively tied to Heretic. Rather than publish a sentiment score built on the wrong subject, we publish nothing here and re-run the scan.
How we verified this
Every listed fact about Heretic — pricing, features, links, status — is re-read from the vendor's own pages by our verification pipeline and each change is logged with a date; the last pass ran on 8 Nov 2026. Analysis is drafted with AI models and checked against our editorial standards before publishing; see how we verify.
Verification history
We have re-verified Heretic 13 times since . Each pass re-reads the vendor's own pages and re-checks every listed field against that evidence; passes where nothing had changed are marked as such.
- — changed domain expiry
- — changed features, our verdict, pricing tiers, description, summary, who should skip it, who it suits, pricing page
- — changed features, who it suits, description, our verdict, hidden costs, integrations, limitations, switching in, switching out, models, who should skip it, pricing tiers, pricing fit, setup time, summary, use cases
- — changed SSL certificate
- — changed platforms, limitations
- — changed link health
Showing the 6 most recent of 13 verification passes.
Free to cite with attribution — this page re-verifies continuously.
How much does Heretic cost in 2026?
Heretic: Free to use. Pricing last verified 8 Nov 2026.
Plans compared
Every published Heretic tier, its price as the vendor publishes it, who it fits, and what it adds over the tier below. The 12-month cost projection follows.
| Plan | Price | Billing | Who it's for |
|---|---|---|---|
| Free | Free | — | Researchers, developers and hobbyists running open-weight models on their own hardware who are comfortable with pip and a terminal.Adds: Starting tier — the whole tool, free software under the AGPLv3-or-later licence, with no paid tier published. |
12-month cost
Project the real annual outlay, including the implied monthly cost when only an annual tier is published.
Vendor list price only. Add-on usage, seat overages, and contract minimums are surfaced under Hidden costs & gotchas.
Where the pricing makes sense
The company stage and team size where Heretic's pricing actually pencils out — and where peers do it cheaper.
Heretic costs nothing to use — it is free software under the AGPLv3-or-later licence. The real spend comparison is against your compute: it sits alongside free self-hosting stacks, and well below the per-seat cost of hosted commercial LLM platforms. Budget for a GPU you control, not a subscription.
Has Heretic's pricing changed?
No plan price for Heretic has moved since we began logging in Aug 2026; the pricing page is re-read on every verification pass.
Who should pick Heretic?
Pick Heretic if you are AI safety researchers studying how refusal and censorship training shapes model behaviour, Developers running local open-weight inference who need unfiltered output from weights they already host, Technical users comfortable with pip, Python environments and command-line model tooling, Open-source contributors and hobbyists who want a maintained alternative to writing abliteration scripts.
Look elsewhere if you are Non-technical users who want a packaged app with a GUI and no Python setup, Organisations that need vendor liability cover, indemnification or a support SLA, Teams in regulated industries requiring documented compliance sign-off on model behaviour, Products that would expose a modified model over a network without accepting AGPLv3 source-disclosure obligations.
Skip it if Skip Heretic if you need a supported commercial product with an SLA, or if you cannot modify and self-host the open-weight model you want to use.
GeneratedReal-world workflow fit
Concrete scenarios for the personas Heretic actually fits — and what changes day-one when you adopt it.
You install Heretic with 'pip install -U heretic-llm', run it against Qwen/Qwen3.5-4B, and then prompt the same benchmark set at the original and patched model to quantify how refusal rate changed.
Outcome: A repeatable before-and-after measurement of refusal behaviour you can write up or hand to collaborators.
Using the Python API, you patch an open-weight model and wire it into a local chat interface that runs on your own machine, with no data leaving your hardware.
Outcome: A local, unfiltered assistant you control, with no per-token bill and no request logs going to a third party.
You generate synthetic completions that require unfiltered model output and script the run through the CLI so it repeats identically across batches.
Outcome: Unfiltered training data produced reproducibly instead of through ad-hoc prompt rewrites.
Use cases
- Run uncensored open-weight models for creative writing and roleplay on your own hardware
- Research how refusal training shapes model outputs by comparing behaviour before and after patching
- Build a private, local chatbot whose responses are not filtered by the model's safety training
- Generate synthetic training data that requires unfiltered completions
- Study or audit the mechanics of censorship patterns in open-weight LLMs
Setup time & first value
How long it actually takes to get something useful out of Heretic — broken out by persona, not the marketing-page minute.
If you already have Python and a model you can run, first value comes within minutes: install with 'pip install -U heretic-llm', then invoke 'heretic Qwen/Qwen3.5-4B' following the site's Tutorial. Most of the elapsed time is downloading model weights and getting your GPU environment working, not the tool itself. Budget an afternoon if this is your first local open-weight model.
Switching to or from Heretic
How to bring data in from common predecessors and how to get it back out — written for the switcher, not the buyer.
- →From manual fine-tuning: replace your training run with 'pip install -U heretic-llm' and a single CLI invocation against the same open-weight checkpoint.
- →From prompt-level jailbreak workarounds: move the effort into a repeatable model patch instead of maintaining fragile prompt templates.
- ↗To a hosted commercial LLM platform: swap the local endpoint for the vendor's API when you need managed uptime and a support contract.
- ↗To a fully custom fine-tune: keep the model but replace the patching step with your own training pipeline if you need tighter control over the modified weights.
Integrations
Resources & guides
Tutorials & Learning
YouTube returned 6 videos for “Heretic”, and we withheld 6: 6 could not be judged, because “Heretic” is a single word that other videos use for other things. We are showing none, because we could not prove any of them are about Heretic.
Official links
What are the best Heretic alternatives?
The closest alternatives to Heretic we list are Gateway and OrcaRouter. Ranked by direct product-type match, not category overlap.
| Alternative | Pricing | Starts at | Last verified |
|---|---|---|---|
| Gateway | Freemium | Pricing not published | — |
| OrcaRouter | Freemium | Pricing not published | — |
| Mindgard | Contact Sales | Custom pricing | 13 Nov 2026 |
Alternatives to Heretic
View all
Gateway
Portkey's AI Gateway gives you one OpenAI-compatible endpoint to route, cache, and govern traffic across 1600+ LLMs and providers.
OrcaRouter
One AI gateway that grades every prompt, routes it across 200+ models at 0% token markup, and enforces guardrails before you're billed.
Popular in AI Governance & Guardrails
Tools that pair well with Heretic
Common stack mates teams adopt alongside Heretic, with the specific reason each pairing earns its keep.
Heretic vs its closest rival
Heretic's closest head-to-head is with Spider Cloud.
Our Heretic vs Spider Cloud verdict written Jul 3, 2026 is more than 90 days old, so it is being re-reviewed against current data and is not shown.
Frequently asked questions
Categories
For makers of Heretic
- Get a Verified Listing for Heretic ($39): a written decision within 1 working day.
- Free listing for another tool: Free reviews are decided in order, within 3 weeks of the review starting.
Used Heretic? Help shape our editorial sentiment research.