A 30B Thinking teacher compressed 50x into a model that fits on a smartwatch.
Available Quantizations
File
Quant
Size
Use Case
qwen3-0.6b-distilled-30b-thinking-sft-f16.gguf
F16
~1.3 GB
Full precision reference
qwen3-0.6b-distilled-30b-thinking-sft-Q8_0.gguf
Q8_0
~700 MB
Near-lossless, desktop/laptop
qwen3-0.6b-distilled-30b-thinking-sft-Q5_K_M.gguf
Q5_K_M
~500 MB
Balanced, mobile
qwen3-0.6b-distilled-30b-thinking-sft-Q4_K_M.gguf
Q4_K_M
~400 MB
Smallest, IoT/edge/smartwatch
Recommended: Q5_K_M for mobile, Q4_K_M for maximum compression.
About the Model
Two-stage build:
Stage 1 — Thinking Teacher Distillation: Qwen3-0.6B distilled from Qwen3-30B-A3B-Thinking on 6,122 STEM chain-of-thought samples. The Thinking variant teacher produces extended reasoning traces with higher-entropy distributions, transferring richer deliberation structure into the student. Proof-weighted cross-entropy (2.5x → 1.5x on derivation tokens) + KL divergence at T=2.0.
Stage 2 — Legal SFT: Supervised fine-tuning on Alignment-Lab-AI/Lawyer-Instruct at conservative learning rate (5e-6) to layer legal reasoning on top of the STEM backbone without overwriting it.
1from llama_cpp import Llama
23llm = Llama(model_path="qwen3-0.6b-distilled-30b-thinking-sft-Q4_K_M.gguf", n_ctx=1024)45output = llm(6"### Instruction:\nProve that the square root of 2 is irrational.\n\n### Response:\n",7 max_tokens=512,8 temperature=0.0,9)10print(output["choices"][0]["text"])
Ollama
bash
1echo'FROM ./qwen3-0.6b-distilled-30b-thinking-sft-Q4_K_M.gguf'> Modelfile
2ollama create stem-legal-tiny -f Modelfile
3ollama run stem-legal-tiny "Explain the difference between a felony and a misdemeanor."
LM Studio
Download any GGUF file from this repo and load directly in LM Studio.
Prompt Formats
STEM derivation (Stage 1):
Solve the following problem carefully and show a rigorous derivation.
Problem:
[Your problem]
Proof:
Instruction-following (Stage 2):
### Instruction:
[Your question]
### Response:
Limitations
0.6B is a hard capacity constraint. The model trades depth for deployability — it will make errors that larger models avoid. Multi-step proofs beyond ~8 steps degrade. Legal reasoning covers general concepts but lacks nuance. Always verify critical outputs. This is not a substitute for formal proof verification, licensed legal counsel, or professional analysis.
Source Model
Full training methodology, hyperparameters, and the two-stage pipeline are documented in:
This is a GGUF-quantized variant. The mathematical foundations (Discrepancy Calculus, Topological Knowledge Distillation) are documented in the source model's card. The discrepancy operator $Df(x)$ and BV decomposition that inform the training pipeline are preserved through quantization — the structural boundaries detected by DISC during training are baked into the weights, not dependent on precision.
This is a GGUF-quantized variant. The mathematical foundations (Discrepancy Calculus, Topological Knowledge Distillation) are documented in the source model's card. The discrepancy operator $Df(x)$ and BV decomposition that inform the training pipeline are preserved through quantization — the structural boundaries detected by DISC during training are baked into the weights, not dependent on precision.
The only BF16 collection in the portfolio. While the broader Convergent Intelligence catalog (43 models, 12,000+ downloads) was trained on CPU at FP32 for $24 total compute, the DistilQwen series was trained on H100 at BF16 with a 30B-parameter teacher. Same methodology, premium hardware. This is what happens when you give the pipeline real compute.
All models use proof-weighted knowledge distillation: 55% cross-entropy with decaying proof weights (2.5× → 1.5×), 45% KL divergence at T=2.0. The proof weight amplifies loss on reasoning-critical tokens, forcing the student to allocate capacity to structural understanding rather than surface-level pattern matching.
