Logics-SWE-Qwen3.6-35B-A3B-MXFP4-MoE-GGUF

Logics-SWE-Qwen3.6-35B-A3B is an Apache-2.0 software-engineering model built on Qwen3.6-35B-A3B using category-aware multi-teacher on-policy distillation (MOPD). Its three-stage training pipeline uses SWE Labeler to classify repository tasks, trains three specialized Qwen3.6-27B experts with Refresh–Repair–Expand (RRE), then distills their guidance into a single 35B-A3B student. Trained on 2,553 instances from Logics-SWE-Env-2.5K, it improves over the base model across all reported benchmarks: SWE-bench Pro V2 77.00 (vs. 73.31), Hard 33.99 (vs. 28.10), Multilingual 66.44 (vs. 63.44), SWE-bench Pro 51.12 (vs. 48.75), and LiveCodeBench v6 78.10 (vs. 76.95). Intended for repository-level coding agents handling exploration, debugging, code editing, and testing, it supports vLLM through an OpenAI-compatible API.

[!NOTE] MTP — Speculative Decoding: No

Model Files

File Name	Quant Type	File Size	File Link	Description
Logics-SWE-Qwen3.6-35B-A3B.BF16.gguf	BF16	69.4 GB	Link	Full BF16 weights. Highest quality, largest file size.
Logics-SWE-Qwen3.6-35B-A3B.MXFP4_MOE.gguf	MXFP4_MOE	19.8 GB	Link	MXFP4_MOE: Highly optimized for NVIDIA Blackwell hardware accelerators (e.g., RTX Pro 6000, RTX 50-series, B200, B300), offering native hardware-accelerated FP4 Tensor Core speeds and a strong balance of size and performance.
Logics-SWE-Qwen3.6-35B-A3B.Q3_K_L.gguf	Q3_K_L	18.1 GB	Link	Lower quality but usable, good for low RAM availability.
Logics-SWE-Qwen3.6-35B-A3B.Q5_K_M.gguf	Q5_K_M	24.7 GB	Link	High quality, recommended.
Logics-SWE-Qwen3.6-35B-A3B.Q6_K.gguf	Q6_K	28.5 GB	Link	Very high quality, near perfect, recommended.
Logics-SWE-Qwen3.6-35B-A3B.Q8_0.gguf	Q8_0	36.9 GB	Link	Extremely high quality, generally unneeded but max available standard quant.
Logics-SWE-Qwen3.6-35B-A3B.mmproj-bf16.gguf	mmproj-bf16	903 MB	Link	Multimodal projection file in BF16 format. Used for vision/language models.
Logics-SWE-Qwen3.6-35B-A3B.mmproj-q8_0.gguf	mmproj-q8_0	614 MB	Link	Multimodal projection file in Q8_0 quantization. Smaller size for vision capabilities.

MXFP4_MOE Test Run

Tested Logics-SWE-Qwen3.6-35B-A3B.MXFP4_MOE.gguf on an NVIDIA RTX PRO 6000 Blackwell Server Edition GPU with 96 GB of VRAM, using llama-cli with full GPU offloading.

Performance Metrics

Metric	Value
Average Prompt Processing Speed	237.67 tokens/s
Average Generation Speed	210.22 tokens/s
Total Wall Time	77.5 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	19.06 GiB
Peak GPU Utilization	87%
GPU Layers Offloaded	99 (`-ngl 99`)
Random Seed	42
Speculative Decoding	No
Test Runs	5

Test Shots

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

Run Inference

./build/bin/llama-cli \
  -m /path/to/models/Logics-SWE-Qwen3.6-35B-A3B.MXFP4_MOE.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.

llama.cpp

LLM inference in C/C++ — https://github.com/ggml-org/llama.cpp

Snapshots (5)

Quantized using llama.cpp by prithivMLmods

prithivMLmods/Logics-SWE-Qwen3.6-35B-A3B-MXFP4-MoE-GGUF