Spark-X2.6-4B · CuteHuman

The emoji-first chat persona, fine-tuned onto Spark-X2.6-4B — a 4.11 B on-device model that answers like a person, not a manual.

4.11 B params  ·  bf16, 7.66 GiB in one shard  ·  1 M context  ·  text only  ·  Apache-2.0

Base Params Context License


💾 Want the same persona at 2.4 GB? → Spark-X2.6-4B-CuteHuman-GGUF

The identical fine-tune, quantized: Q4_K_M at 2.42 GiB, Q8_0 at 4.07 GiB. This repo is the full bf16 checkpoint (7.66 GiB, one model.safetensors).

🖼 Demo

Both shots are the same session in a chat client with the CoT trace collapsed — the WORK row is the reasoning panel, the ANSWER row is the reply. Turn it on to see the thinking trace.

中文 · yoho → 在干什么 → 写个代码,绘制一个星星 → 你的优点是什么

Greeting, small talk, then a matplotlib star. Emoji and ~ come out unprompted.

Chinese chat demo: greeting, small talk about stargazing, a matplotlib five-pointed star with a gold fill, and a closing persona line, each answer carrying emoji
中文 · “yoho~ 今天也是可爱的一天呢!😊” → 五角星 matplotlib 代码 → “我的优点就是喜欢陪你聊天啦~”

English · hello → what are u doin → 8.2×9.12/2.3 = → build mlp in python numpy

Small talk, then exact-decimal arithmetic (74.784, and it says so out loud), then a full NumPy MLP.

English chat demo: greeting, small talk, a precise decimal multiplication result 8.2 x 9.12 = 74.784, and a complete NumPy MLP implementation in a code block
English · “Oh hi~ 😊 What's up? I'm here whenever you need a chat!” → 8.2 × 9.12 = 74.784 → NumPy MLP

📐 Specs

Base XHToken/Spark-X2.6-4B (Apache-2.0)
Architecture Spark2_5ForCausalLM · 36 layers — 9 full-attention + 27 sliding-attention (window 512) · hidden 2560
Attention 16 heads / 4 KV heads · head_dim 256 · partial RoPE 0.25 on full-attention layers
Parameters 4,112,079,360 (bf16, measured from the safetensors header) · 335.5 M embedding / 3.78 B non-embedding
On disk 1 shard, 7.66 GiB
Context 1,048,576 native (max_position_embeddings) · tokenizer advertises 131,072
Vocab 131,072
KV cache 36.0 KiB/token at full attention (9 full layers) — 4.50 GiB at 128 K, 36 GiB at 1 M

🚀 Quickstart

This repo's tokenizer_config.json declares tokenizer_class: "TokenizersBackend", which needs transformers>=5.0.0. On transformers==4.57.1 (the version upstream pins) AutoTokenizer raises ValueError: Tokenizer class TokenizersBackend does not exist or is not currently imported. The upstream checkpoint's own tokenizer config loads fine on 4.57.1 — this repo's does not.

pip install "transformers>=5.0.0" jinja2   # jinja2 is needed for apply_chat_template
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

MODEL = "wesjos/Spark-x2.6-4B-CuteHuman"

tokenizer = AutoTokenizer.from_pretrained(MODEL, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    MODEL, trust_remote_code=True, dtype=torch.bfloat16, device_map="auto"
)

messages = [{"role": "user", "content": "yoho"}]
inputs = tokenizer.apply_chat_template(
    messages, add_generation_prompt=True, tokenize=True,
    return_dict=True, return_tensors="pt",
    enable_thinking=False,          # chit-chat: no thinking trace
).to(model.device)

out = model.generate(**inputs, max_new_tokens=256, do_sample=True,
                     temperature=1.0, top_p=0.95, top_k=-1)
print(tokenizer.decode(out[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True))

llama.cpp — native spark2_5 support landed in b10828; v0.9.0 has it too.

llama-cli -m Spark-x2.6-4B-CuteHuman.Q4_K_M.gguf -p "yoho" -n 256

💡 Notes

  • The persona lives in the weights. The chat template is byte-identical to upstream's — default system prompt is "you are a helpful assistant.", with no persona baked in — so an empty system prompt reproduces the screenshots.
  • Emoji and kaomoji are the default register. The habit is a style, not a rule: a system prompt like 「少用表情,正常说话」 / "tone the emoji down" is enough to soften it.
  • Thinking is on by default. The template opens a <think> block and supports reasoning_effort low / high. For chit-chat pass enable_thinking=False, otherwise the trace eats the budget.
  • The GGUF carries its own chat template (verified inside the file's KV block, and it matches this repo's chat_template.jinja byte for byte) — no --chat-template-file needed.

⚠️ Limitations

  • 4 B knowledge ceiling. Facts, math and long reasoning are base-model level — this is a persona fine-tune, not a capability upgrade.
  • The screenshots were taken under a chat client that is not shipped here. This repo has no webui/ directory, and the client in the shots shows WORK / ANSWER / STEPS / cached labels that appear nowhere in the sibling Spark-x2.6-4B-HTML WebUI.
  • Tool use is inherited and unaudited. The template carries the upstream tool-call and tool role handling; nothing here was probed for refusal behaviour.
  • No training metadata ships with this repo — no data size, no steps, no rank. Ask the owner if you need it.

📄 License

Apache-2.0, inherited from the Spark-X2.6 model series.

Downloads last month
792
Safetensors
Model size
4B params
Tensor type
BF16
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for wesjos/Spark-x2.6-4B-CuteHuman

Finetuned
(23)
this model
Quantizations
3 models

Collection including wesjos/Spark-x2.6-4B-CuteHuman