Domyn Edge Teacher Math (
domyn/Domyn-Edge-Teacher-Math) is a fully open, 4B-parameter dense transformer built by Domyn on top of Domyn Edge SFT to serve as the specialized mathematics domain teacher in its Multi-Teacher On-Policy Distillation (MOPD) post-training pipeline. Released in bfloat16 as part of the second Domyn Edge bundle alongside its training dataset (Domyn-Edge-Teacher-Math-Mix), it was trained via asynchronous Group Relative Policy Optimization (GRPO) on math problems with verifiable final answers—usingmath-verifyto award format-sensitive rewards (1.0 for a properly formatted correct answer, 0.5 for an unformatted correct answer, and 0 otherwise). Intended to score student rollouts token-by-token rather than operate as a general-purpose assistant, it boosts AIME 2026 avg@10 performance by +17.7 points over Domyn Edge SFT (from 61.3% to 79.0%) and outperforms Qwen3.5 4B on HMMT Feb 2025 (76.1%) and Feb 2026 (65.9%). Retaining the exact architecture, tokenizer, 13-language support, and<think>reasoning chat template of Domyn Edge SFT, it is released under an Apache 2.0 license.
| File Name | Quant Type | File Size | File Link | Description |
|---|---|---|---|---|
| Domyn-Edge-Teacher-Math.BF16.gguf | BF16 | 7.87 GB | Link | Full BF16 weights. Highest quality, largest file size. |
| Domyn-Edge-Teacher-Math.Q3_K_L.gguf | Q3_K_L | 2.16 GB | Link | Lower quality but usable, good for low RAM availability. |
| Domyn-Edge-Teacher-Math.Q3_K_M.gguf | Q3_K_M | 2 GB | Link | Low quality, smallest file size. |
| Domyn-Edge-Teacher-Math.Q4_K_M.gguf | Q4_K_M | 2.42 GB | Link | Good quality, default size for most use cases, recommended. |
| Domyn-Edge-Teacher-Math.Q4_K_S.gguf | Q4_K_S | 2.31 GB | Link | Slightly lower quality with more space savings, recommended. |
| Domyn-Edge-Teacher-Math.Q5_K_M.gguf | Q5_K_M | 2.81 GB | Link | High quality, recommended. |
| Domyn-Edge-Teacher-Math.Q5_K_S.gguf | Q5_K_S | 2.75 GB | Link | High quality with slight space savings, recommended. |
| Domyn-Edge-Teacher-Math.Q6_K.gguf | Q6_K | 3.23 GB | Link | Very high quality, near perfect, recommended. |
| Domyn-Edge-Teacher-Math.Q8_0.gguf | Q8_0 | 4.18 GB | Link | Extremely high quality, generally unneeded but max available standard quant. |
Tested Domyn-Edge-Teacher-Math.Q3_K_L.gguf on an NVIDIA RTX PRO 6000 Blackwell Server Edition GPU with 96 GB of VRAM, using llama-cli with full GPU offloading.
| Metric | Value |
|---|---|
| Average Prompt Processing Speed | 1,109.73 tokens/s |
| Average Generation Speed | 267.61 tokens/s |
| Total Wall Time | 56.3 seconds |
| Context Length | 16,384 tokens |
| KV Cache (K) | q8_0 |
| GPU | NVIDIA RTX PRO 6000 Blackwell Server Edition |
| GPU VRAM Capacity | 96 GB |
| Peak GPU VRAM Usage | 4.00 GiB |
| Peak GPU Utilization | 91% |
| GPU Layers Offloaded | 99 (-ngl 99) |
| Random Seed | 42 |
| Speculative Decoding | No |
| Test Runs | 5 |
| Shot | Test |
|---|---|
| Shot 1 | C: Insert Node at End of Singly Linked List |
| Shot 2 | Jogger and Train Passing Time |
| Shot 3 | Two Pipes Filling a Tank |
| Shot 4 | Students in Examination Rooms |
| Shot 5 | Father and Son Ages |
./build/bin/llama-cli \
-m /path/to/models/Domyn-Edge-Teacher-Math.Q3_K_L.gguf \
-ngl 99 \
-c 16384 \
--cache-type-k q8_0 \
--seed 42 \
-n 8192 \
-st \
-p "Your prompt here"
Performance results reflect this specific hardware and inference configuration. Actual throughput may vary depending on the prompt, model, software build, and runtime settings.
LLM inference in C/C++ — https://github.com/ggml-org/llama.cpp
Quantized using llama.cpp by prithivMLmods.
domyn/Domyn-Edge-Teacher-Math