Recommended: Q5_K_M for desktop use, Q4_K_M for mobile/edge.
About the Model
This is a two-stage model:
Stage 1 — DISC-Informed Knowledge Distillation: Qwen3-1.7B distilled from Qwen3-30B-A3B-Instruct on 6,122 STEM chain-of-thought samples using proof-weighted cross-entropy loss (2.5x → 1.5x decay on derivation tokens) and KL divergence at T=2.0. The distillation emphasized multi-step reasoning over final-answer pattern matching.
Stage 2 — Legal SFT: Follow-up supervised fine-tuning on Alignment-Lab-AI/Lawyer-Instruct to add instruction-following capability and legal domain knowledge on top of the STEM reasoning backbone.
The result is a 1.7B model that fits on a phone and can do structured derivations, legal reasoning, and instruction-following.
1./llama-cli -m qwen3-1.7b-distilled-30b-sft-Q4_K_M.gguf \2 -p "### Instruction:\nExplain the doctrine of promissory estoppel and provide a worked example.\n\n### Response:\n"\3 -n 512 --temp 0.0
llama.cpp Python
python
1from llama_cpp import Llama
23llm = Llama(model_path="qwen3-1.7b-distilled-30b-sft-Q4_K_M.gguf", n_ctx=1024)45output = llm(6"### Instruction:\nProve that the sum of two even numbers is even.\n\n### Response:\n",7 max_tokens=512,8 temperature=0.0,9)10print(output["choices"][0]["text"])
Ollama
bash
1# Create a Modelfile2echo'FROM ./qwen3-1.7b-distilled-30b-sft-Q4_K_M.gguf'> Modelfile
3ollama create stem-legal -f Modelfile
4ollama run stem-legal "What is res judicata?"
LM Studio
Download any GGUF file from this repo and load it directly in LM Studio.
Prompt Formats
This model responds to two prompt formats from its two training stages:
STEM derivation (from distillation):
Solve the following problem carefully and show a rigorous derivation.
Problem:
[Your math/physics/engineering problem]
Proof:
Instruction-following (from SFT):
### Instruction:
[Your question or task]
### Response:
Limitations
This is a 1.7B model — it punches above its weight on structured reasoning but has hard limits. It can produce fluent but incorrect derivations. It is not a substitute for formal proof verification, legal counsel, or professional engineering analysis. Verify all outputs independently. Performance is strongest on physics, differential equations, and legal instruction-following. Weaker on underrepresented domains (molecular biology, physiology).
Source Model
Full training details, methodology, hyperparameters, and the DISC-informed distillation approach are documented in the source model card:
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.
