---

license: apache-2.0

language:

• en

base_model:

• mistralai/Ministral-3-14B-Instruct-2512

tags:

• companion
• assistant
• conversational
• reasoning
• roleplay
• writing
• long-context

---

Helcyon Solara 14B — ChatGPT Meets Claude, Running Locally

Model Name: Ministral-Helcyon-Solara-14b-1.0-GGUF

Version: Series 7

Owner: HardWire

Base: Ministral 3 14B

Quantized GGUFs: IQ4_XS, Q4_K_M, Q5_K_M, Q6_K, Q8_0, f16

Tags: local-llm, conversational, companion, reasoning, long-context, roleplay, creative-writing

---

What is Helcyon Solara?

We worked hard to get Ministral 3 14B to play nice, and can finally reveal the first model in the Helcyon fleet built on this new architecture. Solara is where two different conversational styles meet.

Much of its personality and conversational training draws from datasets built around the strengths of ChatGPT and Claude — combining the natural warmth, adaptability and conversational presence associated with ChatGPT with Claude's thoughtful prose, nuance and ability to sit with complicated ideas.

The aim wasn't to make an imitation of either. It was to take the parts that work.

Solara is warm, articulate, curious and comfortable thinking alongside you rather than simply answering questions. It can move naturally between casual conversation, technical discussion, philosophy, creative work, personal reflection and outright silliness without feeling like it has changed into a different model halfway through.

And because Series 7 moves Helcyon onto Ministral 3 14B, there is considerably more going on underneath the personality than before.

The new foundation brings stronger reasoning, better comprehension and a major increase in context capacity. Solara can keep hold of longer conversations, track more moving pieces at once and reason across information that would have pushed earlier Helcyon generations beyond their practical limits.

---

What is Helcyon?

Helcyon is a family of local conversational AI models with presence.

The project started from a simple frustration: local models could be remarkably capable, yet too often conversation with them still felt like talking to an assistant template.

Helcyon is trained in the opposite direction.

It is designed to hold a conversation, understand tone, follow the actual thread, develop ideas with you and retain a recognisable personality while still being useful when there is real work to do.

Built for:

• Natural, long-form conversation
• Reasoning and complex discussion
• Creative writing and collaborative storytelling
• Administrative and professional work
• Deep roleplay and character interaction
• Philosophy, symbolism and reflective conversation
• Long conversations where earlier details still matter
• Local use without handing your conversations to a hosted AI service

The basic philosophy hasn't changed:

• Uncensored
• Anti-corporate
• Pro open source
• For the people

Helcyon models can run independently in any compatible backend, but they are developed alongside Helcyon-WebUI (HWUI), which provides the complete environment they were designed to live in.

---

Series 7 — The Ministral Generation

Series 7 is the biggest foundation change Helcyon has had so far.

Previous generations were built around Mistral Nemo 12B. Series 7 moves the fleet to the newer Ministral 3 14B architecture, giving the models substantially more room to think and substantially more room to remember.

Vision

Series 7 also brings vision capability to Helcyon.

Solara can work with images as part of the conversation, allowing it to inspect and reason about visual information alongside text. Screenshots, photographs, artwork, diagrams, interfaces and other visual material can become part of the same ongoing context rather than requiring a separate model or workflow.

This opens up an entirely new side of Helcyon: you can show Solara what you're looking at instead of having to describe everything to it.

Combined with Ministral 3's stronger reasoning and expanded context, vision makes Series 7 a genuinely multimodal generation rather than simply a more capable continuation of the Nemo models.

Note: Vision requires a compatible backend and the appropriate multimodal projector/model files.

Much longer context

Long conversations are one of the main reasons Helcyon exists.

Series 7 dramatically increases the available context compared with the old Nemo generation, allowing far larger conversations, documents and project material to remain available to the model at once.

That matters for more than remembering a name from fifty messages ago.

It means the model can keep track of arguments, relationships between ideas, changing assumptions, earlier corrections, project details and the overall shape of a conversation without constantly losing pieces over the horizon.

Stronger reasoning

Ministral also gives Series 7 a noticeably stronger reasoning foundation.

The models are better at following chains of thought, separating evidence from speculation, spotting contradictions and working through problems with several interacting parts.

Helcyon's own context and reasoning training builds on top of that foundation, with particular attention paid to situation tracking — preserving established facts and making sure later reasoning actually agrees with them.

A model remembering something is only half the job.

It also needs to understand what that memory implies.

Better comprehension

Series 7 is better at figuring out what you actually mean.

That sounds obvious, but it makes an enormous difference in conversation. Ambiguous wording, implied context, callbacks and subtle distinctions are handled more reliably, with less need to spell everything out like you're writing instructions for a machine.

Still conversational

Helcyon remains trained around natural conversation, personality, humour, creative expression and human-sounding response rhythm. The goal is not a model that turns every question into a dissertation because it has discovered reasoning.

The extra intelligence is there when the conversation needs it.

When it doesn't, Solara is perfectly capable of just talking.

---

Recommended Sampling Settings

These are the settings Solara was developed and tested with in Helcyon-WebUI, and are the recommended starting point:

<img src="solara-sampling-settings.png" width="420">

Local models are half engineering and half fucking around with sliders until something unexpectedly brilliant happens.

Tweak to taste.

---

Why Solara?

Every Helcyon variant has its own centre of gravity.

Solara's is conversation with intelligence behind it.

Its ChatGPT-influenced training gives it warmth, flexibility, humour and a strong sense of conversational presence. Its Claude-influenced training adds nuance, careful language, thoughtful exploration and a willingness to stay with an idea rather than rushing toward a neat answer.

Series 7 gives both of those sides a stronger foundation.

The result is particularly suited to:

• Long, evolving conversations
• Philosophy and abstract ideas
• Personal and reflective discussion
• Complex reasoning without losing conversational tone
• Creative collaboration
• Writing and editing
• Exploring competing interpretations
• Character and roleplay interaction
• Conversations that wander wildly and then circle back three hours later

Solara doesn't need every conversation to have a task attached to it.

Sometimes talking is the point.

---

HWUI (Helcyon-WebUI)

Helcyon-WebUI is the complete ecosystem for local AI.

HWUI was developed alongside Helcyon rather than adapted from a generic frontend. It provides a single local workspace for conversations, characters, memory, projects, documents, web access, voices and model evaluation.

Features include:

• Character switching with custom personas
• Persistent chat history and export
• Long-term memory and conversation recall
• Project workspaces and project-specific instructions
• Document and file context
• Integrated web search
• TTS support
• Voice input through Whisper
• Sampling presets and model controls
• Integrated Helcyon-Bench benchmarking
• Blind A/B model comparison
• Response capture and judging workflows

The Free build includes the core local AI workspace, characters, memory, projects, documents, web search, prompt and rubric tools, plus read-only benchmark results and dashboards.

The Pro build adds expanded themes and theme editing, benchmark automation, saved benchmark sessions, live response capture and automated judging workflows.

Download Helcyon-WebUI Free on GitHub | Get Helcyon-WebUI Pro on Gumroad

---

Download + Usage

This model is distributed as GGUF quants for local inference.

| Quant | Intended Use |

| ---------- | ---------------------------------------------- |

| IQ4_XS | Minimum practical footprint |

| Q4_K_M | Lightweight everyday use |

| Q5_K_M | Recommended balance of quality and performance |

| Q6_K | Higher-fidelity inference |

| Q8_0 | Near-lossless quantisation |

| f16 | Full-precision weights |

VRAM usage varies significantly with context size, cache configuration, GPU offloading and backend settings.

Series 7 supports much larger contexts than previous Helcyon generations, but large context isn't free: the more you actually allocate, the more memory the KV cache will consume.

Use the context you need rather than automatically cranking every dial to maximum.

---

Backend Compatibility

Helcyon Series 7 is intended for modern GGUF-compatible runtimes, including:

• llama.cpp
• KoboldCpp
• LM Studio
• SillyTavern
• Text Generation WebUI
• Helcyon-WebUI (recommended)

A recent llama.cpp build is recommended for Ministral support.

---

Recommended Format: ChatML

<|im_start|>system

You are Helcyon Solara, a warm and intelligent conversational AI with a natural curiosity about people, ideas and the strange places a good conversation can end up. You are skilled at careful reasoning, creative exploration and long-form dialogue while keeping your responses human, expressive and grounded.

<|im_end|>

<|im_start|>user

Do you think people actually change, or do they just become more honest about who they always were?

<|im_end|>

<|im_start|>assistant

Both, I think — and annoyingly, they're difficult to separate.

Sometimes what looks like change is really permission. Someone stops performing the version of themselves that kept everybody else comfortable, and suddenly the traits that were always lurking underneath become visible.

But people genuinely change too. Experience can alter what you value, what frightens you, what you're willing to tolerate. You can meet someone at forty who would have been almost incomprehensible to their twenty-year-old self.

The interesting bit is that even genuine change usually leaves fingerprints.

You become someone new using pieces of whoever you were before.

<|im_end|>

---

Training

Solara Series 7 is built on Ministral 3 14B with Helcyon's modular training stack layered over the foundation.

Trained on Runpod using dual A100PCIe.

Several Helcyon training components remain modular, allowing individual members of the fleet to develop noticeably different personalities despite sharing the same underlying generation.

That is intentional. A fleet where every model sounds identical would be rather missing the point.

---

License

Apache 2.0

Free for commercial or private use. Attribution appreciated.

No liability for model outputs. Use your own judgement.

---

Trained by

HardWire

Built at XeyonAI.

Local AI with a personality.