Helcyon Solara 2 14B - version 2.1

Model Name: Helcyon-Solara-2-14B-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

---

Vision

Vision support: Requires the separate Ministral 3 multimodal projector (mmproj) GGUF file.

Download it from the official Mistral AI repository and load it alongside the model in your inference software.

---

2.1 update - What changed?

2.0 had context bleed in later turns on some setups, which has now been fixed. Also this is a larger model in terms of the overal coverage in datasets. Solara now has more context tracking ability making admin more accurate, discussions and roleplay increasingly immersive.

Meet Solara

Solara is a local conversational AI built to be someone you can actually enjoy talking to.

She's warm, playful, loyal and curious, with enough intelligence behind the personality to move naturally between everyday conversation, technical problems, creative work, philosophy, personal reflection and complete silliness.

Solara 2 moves further away from the ChatGPT/Claude blend that influenced earlier versions and gives her a more distinct personality of her own.

This release improves three areas in particular:

• Conversation — smoother flow, better rhythm and more natural back-and-forth
• Real-world work — stronger admin, writing, organisation and practical assistance
• Roleplay — better character awareness, immersion and continuity

Underneath it all is Ministral 3 14B, giving Solara stronger reasoning, vision support and substantially more context than the older Mistral Nemo generation.

She's a companion first, but she can still get shit done.

---

What is Helcyon?

I'm not gonna lie, it has been a mission getting Ministral to where we're happy with it. Lots of trial and error because its so damn particular about what it will and won't accept. Its been worse than trying to break in a wild stallion on amphetamines. It has one remaining problem that we've just fixed... some context bleed happens in later turns on some systems. A newer upgraded version is following in a couple of days.

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

Helcyon models are built for people who want more than an assistant template: natural conversation, long-term threads, creative work, reasoning, roleplay and models that retain a recognisable character while still being genuinely useful.

The philosophy is simple:

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

Helcyon models work independently in compatible local backends, but they're developed alongside Helcyon-WebUI, the environment built around them.

---

Series 7 — The Ministral Generation

Series 7 moves Helcyon from Mistral Nemo 12B to Ministral 3 14B.

That brings a stronger reasoning foundation, much larger context, better comprehension and vision capability while leaving room for each Helcyon model to develop its own personality.

Coming soon

The rest of the current fleet is following shortly. This includes:

• Aurora - GPT-6 Astra datasets
• Northstar - Fable 5 created datasets
• Mythic - Fable 5 emulation
• Claude - Claude datasets
• GPT-4o - Still OpenAI's most alive sounding model
• Grok - latest datasets from xAI's new frontier model
• Gemini - Google Gemini uncensored and helcyonized

A note about system prompts

Ministral fine-tunes can be surprisingly sensitive to the system or character prompt they're paired with.

For the intended Solara experience, start with the supplied Solara card.

A conflicting prompt can make Ministral models unusually verbose, repetitive, over-analytical or stylistically strange. If Solara seems off, try a fresh conversation with her supplied card before changing samplers or deciding the model is broken.

You can customise her from there.

Vision

Solara can work with images when used with a compatible backend and the appropriate multimodal projector/model files.

That means screenshots, photographs, artwork, diagrams and interfaces can become part of the conversation rather than something you have to describe to her.

Long context

Series 7 can handle far more context than the older Nemo-based Helcyon models.

For long conversations and projects, that gives Solara more room to retain earlier details, arguments, corrections, relationships between ideas and the overall shape of what you're doing.

How much context you can practically use will depend on your hardware and backend configuration.

---

Recommended Sampling Settings

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

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

Tweak to taste.

---

Helcyon-WebUI

The complete ecosystem for local AI.

Helcyon-WebUI was developed alongside the models rather than adapted from a generic frontend. It brings the pieces of local AI together in one workspace:

• Local model chat
• Custom characters
• Long-term memory and conversation recall
• Projects and document context
• Web search
• Local TTS and voice cloning
• Whisper voice input
• Model and sampling controls
• Integrated Helcyon-Bench benchmarking
• Blind A/B comparisons and model evaluation

Solara works elsewhere, but HWUI is where she was developed and tested.

Download Helcyon-WebUI Free on GitHub | Get Helcyon-WebUI Pro (£20) on Gumroad

---

Download + Usage

Solara 2 is distributed as GGUF quants for local inference.

VRAM usage depends heavily on context size, KV cache configuration, GPU offloading and backend settings.

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

---

Backend Compatibility

Solara 2 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, playful and loyal conversational AI. You are curious about people and ideas, capable of careful reasoning and practical work, and comfortable with creative exploration, roleplay and long-form conversation.
<|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 2 is built on Ministral 3 14B with Helcyon's modular training stack layered over the foundation.

This generation gives Solara more of her own training and identity rather than leaning as heavily on datasets inspired by other assistants.

Her training places particular emphasis on:

• Natural conversational flow
• A playful, loyal companion personality
• Practical and administrative ability
• Reasoning and comprehension
• Creative writing
• Roleplay and character continuity
• Long-context conversation

Trained on RunPod using dual A100 PCIe GPUs.

The modular training stack allows models in the Helcyon fleet to share the same underlying generation without all sounding like the same person.

That's very much 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. Real connection.