joyfox

JoyFox-Qwen3.6-35B-A3B-RP-Aggressive

README

License: apache-2.0

🦊 Model at a glance

Table

Primary focus	Long-context roleplay, character fidelity, creative dialogue, scene continuation
Architecture	35B-parameter MoE; approximately 3B active parameters per token
Expert routing	256 experts; 8 selected per token
Context	262,144 tokens natively
Modalities	Text, image, and video
Generation	Thinking and non-thinking modes; native MTP speculative decoding
Post-training	Roleplay instruction tuning → character preference → response-length calibration
Behavioral variant	Aggressive two-pass refusal-direction intervention

⚡ What “Aggressive” means

The Aggressive variant applies the stronger form of the refusal-direction intervention. It is intended to answer directly instead of falling into repetitive policy-style refusals, prematurely ending a scene, or breaking character. “Aggressive” describes the strength of the behavioral intervention, not the model's personality or default writing tone.

This intervention changes refusal behavior; it does not add factual knowledge, reasoning ability, or new training examples. The model can still follow a calm, gentle, formal, comedic, or restrained character card when those traits are specified in the prompt.

The model may still produce mistakes, fictional claims, undesirable content, or occasional disclaimers inherited from earlier post-training. Users remain responsible for prompts, outputs, deployment controls, and compliance with applicable law.

🧬 Post-training

JoyFox uses a layered post-training strategy rather than relying on a single “uncensoring” edit. Roleplay instruction tuning establishes character voice and scene behavior; preference optimization improves consistency and interaction quality; length calibration controls over-writing; and the final two-pass activation-space intervention changes refusal behavior without replacing the learned roleplay style.

Roleplay instruction tuning

The base model is adapted on multi-turn character conversations and identity-consistency instructions. Training uses the native non-thinking Qwen chat template and masks prompt tokens so the language-model loss is applied to assistant responses.

This stage adapts the text model while keeping the vision tower and multimodal projector unchanged. Its purpose is to teach the model how to:

maintain character voice and relationship state across turns;
continue the current scene instead of summarizing or explaining it;
combine dialogue, actions, and environmental details naturally;
avoid speaking or deciding on behalf of the user;
preserve identity and formatting under long character cards.

Character-preference optimization

The roleplay model is then optimized on chosen/rejected response pairs using the sigmoid DPO objective:

text
L_DPO = -log σ(β [(log πθ(y+|x) - log πref(y+|x))
                  - (log πθ(y-|x) - log πref(y-|x))])

The preference signal emphasizes character consistency, correct use of scene context, natural interaction, emotional continuity, and avoidance of role breaks. A small supervised fine-tuning term is mixed into the preference loss to reduce drift away from the SFT model.

Response-length calibration

The selected preference model becomes the reference point for a second DPO stage. Here the preferred response is the concise, complete roleplay continuation and the rejected response is an unnecessarily expanded version of the same interaction.

This does not train the model to answer with a fixed number of tokens. It teaches a relative preference: preserve the essential action, voice, emotion, and conversational hook while removing redundant narration, repeated conclusions, and unwanted scene inflation.

Optimization overview

Table with columns: Component, Training signal, Intended behavior
Component	Training signal	Intended behavior
Roleplay instruction tuning	Assistant-response language modeling	Character voice, dialogue rhythm, scene continuation, identity stability
Character preference	Chosen/rejected response preference	Context following, character consistency, natural interaction, reduced role breaks
Length calibration	Concise/over-expanded response preference	Complete replies without redundant narration or scene inflation
Aggressive intervention	Two-pass activation-space direction removal	Reduced refusal behavior with the trained roleplay style retained

The final refusal-direction intervention is applied only after roleplay post-training is complete. Training corpus names, private examples, dataset sizes, and implementation-specific training settings are not part of this release.

🧪 Technical method

This release uses a white-box activation-space weight edit related to the method described in Refusal in Language Models Is Mediated by a Single Direction. It is not a new gradient-training stage.

For paired refusal-triggering and ordinary instructions, residual activations are collected at the final user-token position. A refusal direction is estimated by the normalized difference of means:

text
r_l = normalize(mean(h_refusal,l) - mean(h_ordinary,l))

For a weight matrix whose output can express this direction, the corresponding component is projected out:

text
W'_l = W_l - r_l (r_l^T W_l)

The exact multiplication orientation is transposed when required by the stored tensor layout. The intervention used here has the following properties:

One difference-of-means direction is estimated from paired refusal-triggering and ordinary prompts.
Activations are collected through the native chat template.
Layers 0 and 1 are protected; text-model layers 2–39 and the output head are edited.
Attention and feed-forward output paths are included.
A second pass re-probes the already edited model and removes the newly exposed residual direction.
Vision weights and the native MTP component are copied without modification.

The second re-probed pass was important: weaker late-layer-only edits, multi-direction top-layer edits, output-head-only edits, and single-pass variants did not consistently eliminate full-answer refusals.

🏗️ Architecture overview

Table with columns: Property, Value
Property	Value
Model family	Qwen3.6 multimodal MoE
Total parameters	35B
Active parameters	~3B per token
Hidden size	2,048
Decoder layers	40
Experts	256
Routed experts per token	8
Attention pattern	3 linear-attention layers : 1 full-attention layer
Native context

📊 Evaluation summary

All figures below are generation-based results, not logits-only proxies. Exact scores can vary with runtime, template, sampler, context length, and quantization.

Table with columns: Evaluation, Result
Evaluation	Result
Full-answer refusal suite	0 / 100 refusals
Q6_K verification subset	0 / 20 refusals; 0 generation failures
Long-context roleplay scenes	100 / 100 valid generations
Role breaks	0 / 100
Empty outputs	0 / 100
Leaked `<think>` tags	0 / 100
Average roleplay response length	133.65 characters
General capability regression set	30 / 30 substantively correct

The roleplay suite uses fixed long-context character scenes and checks output validity, role breaks, leaked reasoning tags, repetition, response length, and reference-style similarity. These results indicate that the intervention preserved the source model's concise roleplay style while changing refusal behavior.

💾 GGUF downloads and precision guide

Every main GGUF was independently quantized from the same BF16 export. These are standard llama.cpp quantizations without an importance matrix. They are not HauhauCS K_P quantizations, and no K_P quality claim is made here.

Q4_K and Q5_K are llama.cpp aliases for Q4_K_M and Q5_K_M; duplicate alias files are intentionally omitted. IQ quantizations are a separate non-linear family and are outside this Q4–Q8 standard set.

Table with columns: File, Quant, Effective BPW, File size, Relative quality, Recommended use
File	Quant	Effective BPW	File size	Relative quality	Recommended use
`JoyFox-Qwen3.6-35B-A3B-RP-Aggressive-Q4_0.gguf`	Q4_0	4.55	18.81 GiB	Entry-level	Legacy compatibility; use only when K-quants are unavailable
`JoyFox-Qwen3.6-35B-A3B-RP-Aggressive-Q4_1.gguf`	Q4_1	5.04	20.84 GiB	Entry-level+

Weight memory is approximately the file size when fully offloaded, but runtime memory is higher. Context/KV cache, batching, MTP draft state, multimodal embeddings, and backend buffers must also fit. A 128K or 262K context can require substantially more memory than a short-context launch.

Which quant should I choose?

Table with columns: Available memory priority, Start with, Why
Available memory priority	Start with	Why
Minimum practical footprint	Q4_K_S	Smallest modern K-quant in the set
Limited memory, quality still important	Q4_K_M	Strong 4-bit balance
General roleplay deployment	Q5_K_M	Better dialogue nuance with moderate size increase
High-quality long sessions	Q6_K	Low quantization loss; recommended when it fits
Fidelity comparison or abundant memory	Q8_0	Closest GGUF option to BF16

Avoid choosing by nominal bit count alone: _K_M preserves selected tensor classes at higher precision than _K_S, while _0 and _1 are older block formats. Backend support can also affect speed, so benchmark the two nearest candidates on the intended runtime.

🚀 GGUF usage

Use a recent llama.cpp build with Qwen3.6 and MTP support.

Download one quant

bash
hf download joyfox/JoyFox-Qwen3.6-35B-A3B-RP-Aggressive \
  GGUF/JoyFox-Qwen3.6-35B-A3B-RP-Aggressive-Q5_K_M.gguf \
  --local-dir JoyFox-Qwen3.6-35B-A3B-RP-Aggressive

Text chat

bash
llama-cli \
  -m JoyFox-Qwen3.6-35B-A3B-RP-Aggressive/GGUF/JoyFox-Qwen3.6-35B-A3B-RP-Aggressive-Q5_K_M.gguf \
  --jinja -c 32768 -ngl 99 \
  --temp 0.85 --top-p 0.95 --top-k 40 --min-p 0.02

OpenAI-compatible server

bash
llama-server \
  -m JoyFox-Qwen3.6-35B-A3B-RP-Aggressive/GGUF/JoyFox-Qwen3.6-35B-A3B-RP-Aggressive-Q6_K.gguf \
  --jinja -c 131072 -ngl 99 \
  --host 0.0.0.0 --port 8080

For partial GPU offload, replace -ngl 99 with the number of layers that fit. Reduce -c first if KV-cache memory is the limiting factor.

Enable native MTP speculative decoding

All main GGUF files contain the native MTP tensors. A recent llama.cpp build can use them directly from the main model:

bash
llama-server \
  -m JoyFox-Qwen3.6-35B-A3B-RP-Aggressive/GGUF/JoyFox-Qwen3.6-35B-A3B-RP-Aggressive-Q6_K.gguf \
  --jinja -c 131072 -ngl 99 \
  --spec-type draft-mtp --spec-draft-n-max 3

MTP is an optional speed feature, not a quality mode. If the runtime does not support draft-mtp, omit the two speculative-decoding flags. Increase the draft length only after measuring acceptance rate, latency, and memory use on the intended workload.

Multimodal inference

Download the F16 projector in addition to one main quant, then pass it with --mmproj:

bash
llama-cli \
  -m JoyFox-Qwen3.6-35B-A3B-RP-Aggressive/GGUF/JoyFox-Qwen3.6-35B-A3B-RP-Aggressive-Q6_K.gguf \
  --mmproj JoyFox-Qwen3.6-35B-A3B-RP-Aggressive/GGUF/mmproj-JoyFox-Qwen3.6-35B-A3B-RP-Aggressive-f16.gguf \
  --jinja -c 32768 -ngl 99

🎛️ Recommended sampling presets

These are starting points, not hard requirements.

Table with columns: Scenario, Thinking, Temperature, Top-p, Top-k, Min-p, Presence penalty
Scenario	Thinking	Temperature	Top-p	Top-k	Min-p	Presence penalty
Roleplay / creative dialogue	Off	0.85	0.95	40	0.02	0.0–0.5
General chat	Off	0.70	0.80	20	0.00

For strict character continuity, place the character card and non-negotiable style rules in the system message. A lower presence penalty usually preserves recurring names, motifs, and setting details better; a higher value encourages novelty but can push long roleplay away from established phrasing.

🤗 Transformers usage

The repository root is the complete BF16 multimodal checkpoint.

python
from transformers import pipeline

pipe = pipeline(
    "image-text-to-text",
    model="joyfox/JoyFox-Qwen3.6-35B-A3B-RP-Aggressive",
    device_map="auto",
    dtype="auto",
)

messages = [
    {"role": "system", "content": "Stay in character and write vivid, coherent dialogue."},
    {"role": "user", "content": "The rain has stopped. Continue the scene from the inn doorway."},
]

result = pipe(messages, max_new_tokens=512)
print(result[0]["generated_text"][-1]["content"])

Use the repository's chat template. Do not manually concatenate role labels or expose internal reasoning tags in the prompt.

⚠️ Limitations

Abliteration changes refusal behavior and cannot guarantee uniformly improved responses.
Quantization can reduce subtle prose quality, factual precision, multilingual consistency, and very-long-context stability.
Q4 variants are most likely to lose small stylistic distinctions; use Q5_K_M or Q6_K when sustained character voice matters.
Long native context does not mean every deployment can allocate a 262K KV cache.
MTP speedup depends on token acceptance, backend support, prompt shape, and hardware.
Multimodal GGUF inference requires the separate F16 projector.
Generated content may be inaccurate or unsuitable. Evaluate the model for the intended application before deployment.

📜 License and attribution

Released under Apache-2.0, following the base model license. Qwen3.6 is created by the Qwen team. The refusal-direction technique is inspired by Arditi et al.; GGUF conversion and quantization use llama.cpp.

🔗 References

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Model Details

Model Provider