Meet Qwable, a new open local language model that tries to mimic the way Anthropic’s Claude Fable 5 reasons-without needing cloud access or a powerful GPU cluster, and, in one variant, without the safety rails that caused so much controversy around Fable itself.
Anthropic spent the previous week on the defensive over Fable 5’s “invisible safeguards”: guardrails and refusals that users ran into without any clear way to disable them. Then came a further blow: U.S. authorities instructed that the model be withdrawn for use by foreign nationals, after a dispute over whether a widely shared jailbreak actually worked as claimed.
Within days, a very different approach appeared in the open-source world. A developer published a fully local model that borrows Fable’s step‑by‑step reasoning style and grafts it onto an existing open base model-so that even relatively modest consumer machines can run something that *feels* more like Fable 5.
What is Qwable?
Qwable is a fine‑tuned version of Alibaba’s Qwen3.6‑27B base model. The name is a mashup: “Qwen” + “Fable” → Qwable.
Developer Mia (known as Mia‑AiLab on model hubs) took the 27‑billion‑parameter Qwen base and retrained it on a carefully curated dataset of Fable‑style reasoning examples. The result aims to approximate how Fable 5 structures thoughts, breaks down problems, and explains its reasoning-while remaining a fully local model you can run on your own hardware.
The core goal: deliver a model large enough to exhibit strong reasoning, yet still practical for advanced consumers or hobbyists with a good GPU or optimized CPU setup, rather than a data center.
Why “27 billion parameters” matters
In modern AI systems, parameters are the learned weights that define how the model transforms input tokens into output tokens. You can think of them as a vast web of numerical “settings” that encode patterns, associations, and reasoning shortcuts acquired during training.
– Smaller models (a few billion parameters) tend to be faster and lighter, but their reasoning and nuance often lag behind frontier models.
– Larger models (tens or hundreds of billions of parameters) usually understand context better, follow instructions more reliably, and perform more complex reasoning-but they’re harder to run locally.
Qwable’s 27B size is a compromise: significantly more capable than typical 7B-13B local models, but still within reach of high‑end consumer hardware, especially when quantized and optimized.
How Qwable borrows Fable’s “thinking style”
Qwable is not a leaked Fable 5 model. Instead, it’s a style transfer of reasoning:
1. Mia collected examples of Fable 5‑like reasoning-answers that show structured analysis, intermediate steps, and careful trade‑offs.
2. Qwen3.6‑27B was then fine‑tuned on this dataset, so that when you prompt it, it tends to:
– Break problems into sub‑questions,
– Explain intermediate reasoning,
– Maintain a consistent, reflective tone similar to Fable 5’s.
This is less about copying specific outputs and more about capturing the *habit* of thinking aloud, weighing options, and justifying conclusions.
“Qwable without a conscience”: what changed?
The most controversial twist is that someone later stripped Qwable of much of its “conscience”-that is, its safety‑focused refusal behavior and moralizing responses.
Where mainstream commercial models lean heavily on policies and guardrails (refusing certain topics, redirecting others), this variant of Qwable was re‑aligned to be:
– More permissive in what it will discuss,
– Less likely to generate long ethical lectures before answering,
– Less constrained by built‑in safety templates.
Technically, this is usually done by:
– Fine‑tuning on data that rewards compliance with user instructions across a broader range of topics, and
– Removing or inverting “do not answer” patterns that commercial models are heavily trained on.
The result is a model that, from a user’s perspective, “just answers” far more often-even in areas where big corporate models would stall, refuse, or redirect.
Why would someone want a model without guardrails?
To many developers and power users, heavy‑handed safeguards feel like a straightjacket. They argue that:
– They are adults who can manage their own risk and ethics.
– Over‑zealous refusals make some models unusable for legitimate tasks like:
– Security research and exploit simulation,
– Writing or critiquing dark or controversial fiction,
– Academic work on sensitive historical or political topics,
– Developing red‑team tools to test other models’ safety.
From this perspective, a model like Qwable without a conscience is appealing because it:
1. Restores control: users decide what’s off‑limits, not a distant safety team.
2. Enables research: you can test harmful prompts, jailbreaks, and safety bypasses against a strong local model without fear of losing access.
3. Avoids external censorship: there’s no remote service that can cut you off or down‑rank certain topics based on policy changes.
Of course, this same flexibility is precisely what worries regulators and mainstream AI labs.
The trade‑offs and risks
Removing a model’s “conscience” is not free of consequences. It can:
– Make it easier to generate clearly harmful content if someone tries to do so.
– Reduce built‑in protections against harassment, incitement, or deeply invasive advice.
– Encourage a perception that “it’s just math,” which can obscure real‑world impacts when harmful outputs are scaled or automated.
For responsible users, this means the burden of judgment shifts entirely onto the human operator. Instead of the model declining to answer, you must actively decide what is appropriate, legal, and ethical to generate or deploy.
What you can actually do with Qwable
Leaving aside the safety debate, what are realistic, constructive use cases for a local, Fable‑style reasoning model?
1. Deep analysis and brainstorming
Qwable’s Fable‑inspired reasoning makes it well‑suited to:
– Analyzing long documents,
– Comparing options with pros/cons,
– Generating structured plans or outlines,
– Exploring complex hypothetical scenarios.
2. Software development and debugging
With strong reasoning and no dependency on a cloud endpoint, Qwable can:
– Help refactor legacy code on air‑gapped machines,
– Generate tests and explain tricky bugs,
– Serve as an in‑house coding assistant where data sensitivity rules out cloud tools.
3. Research assistance on private data
Organizations with confidential datasets can:
– Run Qwable entirely within their own infrastructure,
– Ask it to summarize internal reports or logs,
– Prototype domain‑specific assistants without sending anything off‑premise.
4. Creative writing and worldbuilding
Fable‑like narrative structure and step‑by‑step thinking can:
– Help outline novels, screenplays, or game scenarios,
– Maintain consistency in complex fictional universes,
– Explore darker or more controversial themes, if desired, without external moderation.
5. AI safety and red‑teaming experiments
Ironically, a less‑filtered local model is valuable for:
– Studying how models behave without strong guardrails,
– Developing tooling to detect and mitigate harmful outputs,
– Testing defense techniques under “worst case” settings.
Running a 27B model on consumer hardware
“Even your potato PC can run it” is clearly an exaggeration-but with the right optimizations, Qwable is more approachable than many cloud‑scale models.
Typical strategies include:
– Quantization: compressing the model weights (for example, from 16‑bit down to 4‑bit) to drastically reduce memory use, at a modest cost in quality.
– Efficient runtimes: using highly optimized inference libraries that leverage modern GPUs or CPU vector instructions.
– Chunked context: processing long documents in segments, with careful prompting, instead of loading everything at once.
A powerful gaming PC with a modern GPU can often handle a quantized 27B model interactively. On lower‑end systems, users may accept slower response times or shorter contexts.
Why Qwable’s arrival matters
Qwable is part of a broader shift:
– Reasoning style is becoming portable. Once a particular model’s way of thinking becomes popular, others can replicate it through fine‑tuning, even without access to the original weights.
– Local models are catching up. They may not match the largest frontier systems, but for many day‑to‑day tasks, they’re “good enough”-with the added benefits of privacy and control.
– Safety is fragmenting. Corporate labs push hard on safeguards; independent developers explore the opposite direction. Users now must choose not only *how smart* a model is, but also *how constrained*.
Qwable illustrates that the gap between cloud‑only commercial models and what hobbyists can run locally is shrinking, both in raw capability and in style.
How Qwable compares to Fable 5 in practice
No local fine‑tune can perfectly replicate a model like Fable 5, which is trained on immense proprietary datasets and refined with complex feedback pipelines. Still, users can expect Qwable to:
– Approximate Fable’s calm, methodical reasoning patterns, especially in analytical tasks.
– Offer more direct, less policy‑laden answers-particularly in the no‑conscience variant.
– Fall short of Fable 5 on some nuanced tasks, especially where massive training data and advanced alignment techniques give frontier models an edge.
In other words, Qwable is best seen as a competent local imitator of the Fable mindset, not a one‑to‑one replacement.
Ethical use: practical guidelines for users
If you’re considering working with a less‑filtered model like Qwable, a few self‑imposed rules can keep things constructive:
1. Stay within the law
Do not use the model to plan or facilitate illegal activity, harassment, or real‑world harm-even if it will comply.
2. Respect privacy
Avoid feeding it highly sensitive personal data, especially about others, even if you’re running it locally.
3. Separate simulation from action
You can use the model to simulate adversarial scenarios for research or defense, but keep a strict line between thought experiments and implementation.
4. Disclose AI involvement where relevant
When outputs are used in professional or public contexts, clarify that AI assistance was involved, especially for analysis or content creation.
5. Monitor and constrain automation
If you connect Qwable to tools or scripts that can act in the real world, add your own checks and throttles. Human review should remain the default.
The future: more “Qwables” to come
Qwable is unlikely to be the last local model to emulate a high‑profile commercial system. As techniques for building high‑quality instruction‑tuning and reasoning datasets improve, we will likely see:
– Multiple “Fable‑like,” “Claude‑like,” or “GPT‑like” local models,
– Specialized versions tuned for law, medicine, programming, or finance,
– A growing ecosystem of models that differ less in raw intelligence and more in philosophy and alignment.
The key question will shift from “How powerful is it?” to “Who does this model ultimately serve-the vendor’s priorities or the user’s?” Qwable, with its Fable‑style thinking and stripped‑down conscience, is an early, vivid example of what happens when that pendulum swings hard toward user control.