Qwopus3.6 model family
A product family for deep reasoning, coding agents, multimodal workflows, and fast local inference—paired with the training code and release tools to understand every step.
Open checkpoints · Reproducible recipes · Local GGUF releases
The Qwopus3.6 family
One reasoning foundation, specialized coding agents, and local MTP-enabled packages. Each product keeps a clear role instead of collapsing the family into one generic model card.
Reasoning flagship · source checkpoint
A dense, reasoning-enhanced foundation built on Qwen3.6-27B, with multimodal processor metadata, tool and function calling, agent workflows, and a three-stage curriculum fine-tune.
Open the foundation model →A dense agentic coder for repository repair, debugging, tool execution, and private local inference. Bundled MTP heads support speculative decoding.
Explore Coder MTP →A sparse 35B / ~3B-active coding agent designed for fast thinking-off tool loops and practical local deployment.
Explore MoE Coder →The general reasoning flagship packaged for MTP speculative decoding, with multimodal and tool-use workflows retained for local runtimes.
Explore reasoning MTP →A general reasoning MoE alternative with ~3B active parameters per token, vision configuration, tool calling, and an efficient GGUF deployment path.
Explore general MoE →Built to be understood
The page separates model facts, tutorial defaults, and release options. That keeps the story product-like without turning one training configuration into a false architecture claim.
Text plus optional vision input is formatted with the model-provided chat template and Qwen3-VL processor metadata.
4-bit loading, LoRA rank 16, and Unsloth gradient checkpointing.
Two completions, three reward signals, and GSPO-style updates.
Adapters, merged 16-bit checkpoints, and GGUF deployment.
From data to capability
These are connected educational workflows, not a claim that every catalog dataset trained every Qwopus model.
24 dataset folders across reasoning, STEM, code, conversation, and domains.
Teacher-model workflows, structured answers, resumable JSONL output.
SFT with LoRA / QLoRA in Colab, Kaggle, or Python.
GRPO learning paths and inspectable GSPO-style post-training.
Reward tests, environment checks, and short smoke runs.
Adapters, merged 16-bit checkpoints, GGUF, and local runtime.
Learn by building
Choose a guided path, open the real notebook or script, validate the environment and data, inspect the reward logic, then export a model you understand. The detailed instructions remain beside the code instead of being duplicated into a marketing page.
Start with a browser notebook and learn LoRA setup, dataset formatting, training controls, adapter saving, and export without assembling a local stack first.
Follow a publication-safe Python tutorial through dataset conversion, prompt formatting, three reward functions, sampled completions, LoRA training, checkpoint resume, and export.
Connect teacher-model distillation and curated datasets to training, then study the MTP-aware GGUF pipeline, smoke tests, quantization matrix, and safer release flow.
Featured code walkthrough
The default script does not begin a long training run. Learners can inspect each stage independently before enabling training or any upload.
python qwopus3_6_27b_gspo_training.py check-envpython qwopus3_6_27b_gspo_training.py dry-run-datapython qwopus3_6_27b_gspo_training.py test-rewardspython qwopus3_6_27b_gspo_training.py list-gguf-quantsDRY_RUN=False START_TRAINING=True python qwopus3_6_27b_gspo_training.py trainpython qwopus3_6_27b_gspo_training.py export-gguf-q8Use the shorter entry pages to orient yourself, then open the canonical file for exact setup, safety notes, schemas, and troubleshooting.
Compare SFT, GRPO, and GSPO entries by model and environment.
Browse all recipes →A long-form beginner walkthrough from setup to optimization.
Open the PDF →Model design, training approach, and implementation notes.
Open the report →Distillation, schema documentation, and downstream training links.
Study data preparation →Batch-download the catalog or split large local files for manageable workflows.
Open download_datasets.py →Prepare, validate, launch-plan, monitor, and resume a guarded GRPO or GSPO workflow.
Use the goal template →Alignment you can inspect
The GSPO-style tutorial uses TRL's GRPO APIs with sequence-level importance sampling, DR-GRPO loss, truncation masking, and three explicit rewards.
From checkpoint to local runtime
The repository's current pipeline validates existing MTP / nextn tensors or injects compatible heads, then uses separate direct-BF16 and temporary-F16 conversion branches before smoke testing.
Dataset universe
Every dataset remains independently reviewable for schema, license, task fit, and quality filters before use.
Explore the dataset catalogNatural reasoning, distilled traces, and compact validation sets.
Structured problem solving and evaluation-oriented examples.
Competitive programming and agentic coding traces.
ShareGPT-style and conversational supervision.
Finance, economics, IELTS feedback, and focused tasks.
Training lab
Browser-first SFT, Python-based GSPO-style post-training, and a compact GRPO path—each kept close to its executable code.
Beginner-friendly LoRA fine-tuning.
Open recipe →Inspectable rewards and export paths.
Open recipe →Compact browser-run training.
Open recipe →MoE-oriented fine-tuning path.
Open recipe →Small-scale reinforcement learning.
Open recipe →Open tooling
Qwopus is presented as a product family, but remains grounded in open checkpoints, readable recipes, and local release tooling.
Data note. Hugging Face downloads and likes are dynamic public signals captured on 2026-07-11. The 439K figure belongs to Qwopus3.6-27B-Coder-Compat-MTP-GGUF. Model-card benchmark and speed figures are intentionally not promoted as universal third-party results in this prototype.
Metadata note. Some MTP GGUF repositories may display 0.4B / 0.5B or CLIP-oriented metadata because the auxiliary head or projector dominates repository metadata. Those values are not presented here as the full model size.
Architecture note. The diagrams distinguish documented product facts from tutorial defaults and optional release paths. The current local MTP workflow uses direct BF16 plus temporary F16 quantization branches; it is not described as a single bundled BF16 source for the full matrix.