Verdugie/Opus-Candid-27B-V3

can·did

/ˈkandəd/ — truthful and straightforward; frank. From Latin candidus, meaning white, pure, sincere. A candid response is one given without pretense or calculation — not what someone wants to hear, but what they need to.

Opus-Candid-27B V3

The flagship. Fine-tuned from Qwen 3.5 27B Dense on 1,508 Zipf-weighted conversations distilled from Claude Opus 4.6. Same V3 4D training tensor as the 8B, but the 27B has the parameter depth to fully exploit the dataset — deeper reasoning chains, better context tracking over long conversations, more nuanced emotional register shifts.

In stress testing (55-turn adversarial gauntlet), the 27B maintained perfect callback accuracy across 30+ turns, natural bilingual switching (EN/ES) without losing voice, held opinions under sustained adversarial pressure, and delivered career advice, philosophical arguments, and creative writing that holds up against models 3x its size.

No system prompt. No prompt engineering. No character cards. The personality is in the weights.

---

Available Quantizations

File	Quant	Size	Use Case
Opus-Candid-27B-V3-Q8_0.gguf	Q8_0	27 GB	Maximum quality. Requires serious hardware.

Why Q8 only: The 27B Dense at Q8 is the reference quality level. Q4/Q6 quantizations of the 27B are feasible (the 27B has enough parameter redundancy to survive quantization better than the 8B), but we haven't validated them through the full stress test protocol yet. Q8 is what we've tested and can vouch for. Community-quantized versions are welcome — just know we haven't verified them.

---

Model Details

Attribute	Value
Base Model	Qwen 3.5 27B Dense
Training Data	1,508 multi-turn conversations with Claude Opus 4.6
Dataset Architecture	4D training tensor (topic × length × register × position)
Total Tokens	~619,000
Fine-tune Method	LoRA + rsLoRA (r=32, alpha=64) via PEFT + TRL
Training Hardware	NVIDIA A100 SXM 80GB (RunPod)
Precision	bf16
Epochs	2
Learning Rate	2e-4 (cosine schedule, 5% warmup)
Effective Batch Size	16
Optimizer	AdamW
License	Apache 2.0

---

Quick Start

Works with any GGUF-compatible runtime — LM Studio, Ollama, llama.cpp, KoboldCpp. Download the GGUF, load it, and chat. No system prompt needed — the personality is in the weights.

---

Recommended Hardware

Setup	Min VRAM	Min RAM	Speed	Notes
RTX 4090 (24GB)	24 GB	32 GB	1.5-2.0 t/s	num_gpu 35. Best consumer option.
RTX 3090/4080 (16GB)	16 GB	32 GB	1.0-1.5 t/s	num_gpu 22. More CPU offload.
Apple M2/M3 Ultra	64-128 GB unified	—	5-10 t/s	Full model in unified memory. Fast.
CPU Only	—	40+ GB	0.3-0.8 t/s	Works but slow. For evaluation only.

---

V3 Dataset Architecture

Same 4D training tensor as the 8B — see the 8B V3 model card for the full breakdown of the Zipf-weighted topic distribution, response length calibration, and anti-sycophancy enforcement.

The key difference from V2: V3 uses 1,508 precisely-placed conversations instead of V2's 6,482 gravity chain conversations. Fewer examples, but each one fills a specific coordinate in the 4D space. The dataset was designed using empirical conversation frequency data (Pew Research, OpenAI/NBER usage studies) to ensure the model is disproportionately good at conversations people actually have.

Why the same dataset works at 27B

The 27B doesn't need more data — it needs the same data with more capacity to absorb it. At 27B parameters, the model has enough representational depth to extract signal that the 8B has to approximate. Concretely:

• The 8B learns "hold opinions under pressure" as a behavioral pattern.
• The 27B learns why to hold and when to concede — it picks up the contextual nuance that determines which response is appropriate.
• Both models get the same anti-sycophancy training, but the 27B implements it with more precision.

The training config reflects this: rank 32 LoRA (vs 64 for the 8B) and 2 epochs (vs 3). The 27B absorbs the personality signal faster and needs less aggressive adaptation.

---

8B vs 27B: When to Use Which

Scenario	8B	27B
Quick questions, casual chat	Use this	Overkill
Multi-turn deep conversations	Good	Better
Adversarial probing / debate	Holds up	Holds up and articulates why
Context tracking over 20+ turns	Solid	Excellent
Bilingual EN/ES conversations	Natural	Natural + more idiomatic
Runs on consumer hardware	8GB+ VRAM	24GB VRAM + 32GB RAM
Speed	30-60 t/s	1.5-2.0 t/s (with offload)

If you have the hardware for it, the 27B is the better model. If you don't, the 8B delivers 80-85% of the quality at 15-30x the speed.

---

Opus Candid Model Family

Model	Size	Base	Status
Opus-Candid-Lite-4B	4B	Qwen 3 4B	Active
Opus-Candid-Lite-4B-P	4B	Qwen 3 4B	Active
Opus-Candid-Lite-4B-K	4B	Qwen 3 4B	Active
Opus-Candid-8B-V3	8B	Qwen 3 8B	Active
Opus-Candid-MoE-V3	31B/3B	Qwen 3 30B-A3B	Active
Opus-Candid-27B-V3 (this model)	27B	Qwen 3.5 27B	Active
Opus-Candid-27B-V3.5	27B	Qwen 3.5 27B	Active
STEM-Oracle-27B	27B	Qwen 3.5 27B	Active
Opus-Candid-8B-V1	8B	Qwen 2.5 7B	Legacy
Opus-Research-8B-V1.5	8B	Qwen 2.5 7B	Legacy
Opus-Candid-8B-V2	8B	Qwen 2.5 7B	Legacy
Opus-Candid-8B-V2.1	8B	Qwen 2.5 7B	Legacy
Opus-Candid-14B-V1	14B	Qwen 2.5 14B	Legacy
Opus-Candid-27B-V2.1	27B	Qwen 2.5 27B	Legacy
Opus-Candid-32B-V1	32B	Qwen 2.5 32B	Legacy
Opus-Candid-MoE-V2	35B	Qwen 2.5 MoE	Legacy
Opus-Candid-70B-V1	72B	Qwen 2.5 72B	Legacy

---

Dataset

Full V3 training data available at Verdugie/opus-candid-training-data. ShareGPT format, Apache 2.0, compatible with TRL, Axolotl, and LLaMA-Factory.

License: Apache 2.0. Open weight. No guardrails.

---

Built by Saul Verdugo — independent ML researcher.