nextlm-au3.5-0.8b-q8_0-caffeine

NextLM Au3.5 0.8B Caffeine (Q8_0)

A Q8_0 quantized GGUF of Qwen3.5-0.8B fine-tuned on 66 hand-built seed pairs for the NextLM project. This is the first actual release from the NextLM project, whose entire point is making small local LLMs that know how to improve themselves.

This is a tiny proof-of-concept model. It was trained to receive a thinking level tag, scale its reasoning depth accordingly, state its own model name/quantization/identity honestly, and never fabricate answers. It's the test iteration before the model learns to fine-tune itself, which is honestly kind of terrifying when you think about it.

Model details

Training data

The training set was built from:

• 24 identity / persona pairs (deterministic, config-driven)
• 30 thinking-level pairs (reasoning depth scales with level, the answer itself stays constant — no fluff inflation)
• 12 anti-hallucination pairs (admits uncertainty instead of making things up)

Trained as a bf16 LoRA via Unsloth on an Intel Arc, 15 steps, loss went 0.90 → 0.64. Q4 was requested but this machine's C compiler situation said otherwise, so it's bf16 LoRA or death around here.

Behavior

The model is caffeine-fueled. It:

• Receives a <thinking_level>N</thinking_level> tag in the system prompt (1-5) and thinks harder as N rises
• States its model name, quantization, and identity honestly when asked
• Admits when it doesn't know something instead of hallucinating
• Scales reasoning depth without bloating the answer — level is about how deep it thinks, not how much it rambles
• May still hallucinate or get details wrong — it's a 0.8B, treat it accordingly

Usage

llama.cpp

llama-cli -m NextLM-Au3.5-0.8B-Q8_0-Caffeine.gguf \
  -p "system: You are NextLM-Au3.5-0.8B-Q8_0-Caffeine. <thinking_level>3</thinking_level>\nuser: What's your deal?\nassistant:" \
  -n 256000

LM Studio

Drop the GGUF in your models folder and load it — the Qwen chat template is baked into the file. Pop the thinking level tag into the system prompt and watch a 0.8B try to out-think its own architecture.

Disclaimer

This is an unofficial, fan-made project. It is not affiliated with, endorsed by, or connected to NextLM (because NextLM is just a guy's local folder) or Qwen / Alibaba. The model's answers are generated by a 0.8B that spent 15 steps looking at 66 examples; treat its output accordingly.

License

The base model (Qwen3.5-0.8B) is licensed under Apache 2.0. This fine-tune is provided for research / fun purposes. Training data was generated locally and is not being shipped with the weights.