HeatSeeker
Imagination with a lock-on.
Point it at an idea. Give it a little room. See where the trail goes.
DeepSeek V4 Flash 0731 284B A13B MoE LoRA Fine-Tune NSFW Creative Writing Roleplay No Boundaries SEEK → CHARACTER → IMAGE → TENSION → MOMENTUM → SURPRISE → LOCK ◉ HEATSEEKER // FLIGHT TEST creative payload // target locked // no limits Primary Mission Find what you're after.HeatSeeker is built for your writing. Your prompt gives it the target but not a map. HeatSeeker commits, improvises, develops, and keeps chasing the signal instead of flattening everything into safe, generic prose.
Guidance Intent → Trajectory Payload Character + Voice + Momentum Target The Spark 🔥 So... what is HeatSeeker? start hereHeatSeeker is a creative-writing fine-tune of DeepSeek V4 Flash 0731. We've trained the base model to be lively, imaginative, and more eager to turn every request into a story.
HeatSeeker is better company inside an imaginary space: more willing to choose a direction, inhabit a voice, develop a scene, follow an emotional current, and pursue what your imagination really reaching for.
The metaphor: a heat-seeking missile does not need you to manually steer every inch of its path. You give it the target; the guidance system does the rest. HeatSeeker aims for that feeling in creative writing.🎯 Intent Tracking
Follow the imaginative destination behind the prompt, not just its literal wording.
🔥 Creative Commitment
Choose details, make moves, establish texture, and stop hovering outside the scene.
🎭 Character Presence
Keep people sounding like people instead of flattening everyone into one assistant voice.
🚀 Narrative Momentum
Push the scene somewhere. Escalate. Complicate. Reveal. Surprise. Resolve.
✨ What kind of writing is it chasing? the fun bitHeatSeeker is not tuned for a single genre or style. HeatSeeker is creative energy with direction.
🌒 Atmosphere
Scenes with mood, sensory detail, implication, and a sense that the world extends past the paragraph.
💬 Dialogue
Distinct voices, subtext, interruption, friction, rhythm, and people who do not all sound equally polite.
🌀 Weirdness
Enough willingness to take the strange option when the strange option is the more interesting one.
🎬 Momentum
Less standing around admiring the setup. More things actually happening because of it.
🧠 Continuity
Carry forward character motives, established details, and the emotional logic of prior turns.
✨ Novelty
Prefer the specific, peculiar, memorable image over the stock phrase waiting nearest the surface.
The locked-on feelingYou toss in a premise that has some heat in it, and instead of asking you to do all the steering, the model catches the trail.
🛰️ How do I get a good lock? promptingGive HeatSeeker a creative vector: a mood, an image, a character problem, a dramatic contradiction, a voice, or a destination. You do not need to micromanage every detail.
Example
A lightweight template for giving HeatSeeker a clear narrative target. Copy, customise, and drop it directly into your system prompt.
## CURRENT STORY PARAMETERS
**Premise:**
[Describe the situation or world.]
**User role / perspective:**
[Who the user controls or how they participate.]
**Characters:**
[List important characters and defining traits.]
**Setting:**
[Place, period, atmosphere, relevant world rules.]
**Tone:**
[e.g. intimate psychological drama, pulpy adventure, gothic horror, romantic comedy.]
**Narrative priorities:**
[What you especially want the story to explore.]
**Important constraints:**
[Any facts, stylistic requirements, character boundaries, or structural rules that must remain consistent.]
**Opening situation:**
[Where the story begins and what is already in motion.]
🧪 Fine, show me the nerd stuff.
training
All the serious bits are still here. They are simply not allowed to stand in the foyer wearing a clipboard.
284B Base model total parameters 13B Base parameters active per token 4,546 Completed training steps 264 Validation examples| Item | Value |
|---|---|
| Base model | abliterated DeepSeek-V4-Flash-0731 |
| Architecture | 284B A13B Mixture-of-Experts (MoE) |
| Fine-tuning method | LoRA |
| Primary target | Creative conversation / imaginative prose / NSFW roleplay |
| Training steps | 4,546 |
| Validation examples | 264 |
| Training tokens | ~27.5M |
A curated conversational corpus selected and cleaned for imaginative dialogue and creative-writing behaviour.
Conversation-Level Cleaning
Bad threads were candidates for removal as whole conversations, not merely patched line-by-line.
Structural Repair
Punctuation, quotation structure, spelling, formatting artefacts, and malformed text were reviewed and repaired.
Human Adjudication
Ambiguous samples and edge cases received manual review rather than being trusted blindly to automation.
Multi-Turn Shape
The training material preserves conversational development instead of reducing everything to isolated prompt-response pairs.
🎛️ What knobs should I turn? sampling Off / Low Reasoning Effort 0.8-1 Temperature 0.95 Top-P 0.05 Min-P 1.1 Repeat Penalty If you want stranger prose, increase temperatute carefully. If you want tighter continuity, bring it back down. HeatSeeker should have room to move, not enough room to leave the postcode. 🚀 Small launch pad? We put an ICBM on a Toyota Hilux. minimum hardware testHeat Seeker is a 284-billion-parameter model. That sounds like something you launch from a reinforced concrete bunker with its own electrical substation. But, at only 13-billion active parameters, it's faster than you might expect.
So naturally, we put it on a laptop.
RTX 4070M GPU
8 GB VRAM. Seriously. 128 GB System RAM Thinking Enabled ❌ Draft Model Disabled ~4.5 t/s And away she goes Target acquired. At roughly 4.5 tokens per second, it's fast enough to sit down, start a story, and actually enjoy the conversation.🚚 The Launcher
Laptop-class hardware: an RTX 4070 Mobile with 8 GB VRAM, 128 GB DDR5-5200 system memory, and just two model layers offloaded to the GPU.
💣 The Payload
The IQ2_XS Heat Seeker GGUF, weighing in at approximately 88.9 GB. A 284B MoE with about 13B parameters active per token.
🔥 The Result
About 4.5 tokens per second in ordinary generation. Not "technically running." Not "go make coffee between tokens." Actually usable creative-chat speed.
🛩 Assisted Flight Control
This run used no speculative draft model.
📦 Which warhead—sorry, build—do I download? files| Build | Link | Notes |
|---|---|---|
| LoRA adapter | LINK | Optional adapter-only release |
| GGUF | LINK | IQ1_M |
| GGUF | LINK | IQ2_XS |
| GGUF | LINK | IQ3_M |
| GGUF | LINK | Q4_K_M |
| Evaluation | Result |
|---|---|
| Locks on any target (no refusals) | ✅ |
| NSFW roleplay | ✅ |
| Creative, captivating, coherent responses | ✅ |
| Multi-turn character consistency | ✅ |
| Long-form narrative coherence | ✅ |
- This is a specialised creative-writing fine-tune, not an attempt to improve every capability of the base model.
- Creative outputs are stochastic and can change substantially with sampling and prompt framing.
- This model is not intended or evaluated for agentic capability. USE AGENTICALLY AT YOUR OWN RISK.
HeatSeeker is derived from DeepSeek V4 Flash 0731. The upstream model is distributed under the MIT License.
If merged weights are redistributed, preserve the applicable upstream DeepSeek copyright and MIT licence notice. Your own fine-tuning and modification credit can be listed separately.
| Contribution | Credit |
|---|---|
| Base model | DeepSeek — DeepSeek V4 Flash 0731 |
| Post-training merged base | Huihui-DeepSeek-V4-Flash-0731-abliterated-GGUF|
| Fine-tuning / curation | UltimateIntent |
| Heat Seeker release | August 2026 |
See the full upstream MIT licence notice in the repository
LICENSE file.
- LampLighter — training guidance / tooling.
- Morrow — conversion / quantisation tooling.
- UltimateIntent — dataset generation / curation and GPUs.