Colibro/Qwen-AgentWorld-35B-A3B-INT4

Qwen-AgentWorld-35B Colibro INT4

Colibro-format INT4 sparse-MoE checkpoint converted from Qwen/Qwen-AgentWorld-35B-A3B.

Intended for the Colibro / qwen.exe runtime (OpenAI- and Anthropic-compatible APIs, web UI, CUDA expert tiering). This is not a drop-in replacement for vanilla Hugging Face Transformers inference.

License

Apache 2.0 (inherited from the base Qwen-AgentWorld-35B-A3B weights). See the base model LICENSE.

Format

• Layout: Colibro MoE shards named colibro-NNNNN.safetensors
• Size: ~16.8 GB total (21 weight shards + tokenizer / config sidecars)
• Architecture: Qwen3.5-family sparse MoE text — 40 layers, 256 experts, 8 experts/token, hidden size 2048 (model_type: qwen3_5_moe_text)

Precision policy

Component	Precision
Dense / routed expert matrices	INT4
Embedding / output matrices	INT8 (when present as I/O tensors)
Norms, routing, other small tensors	FP32

Conversion note (MTP / vision)

The source AgentWorld tree did not include MTP or vision weight keys at conversion time (mtp=0, vision=0 in the converter report). The published checkpoint is therefore a text INT4 Colibro MoE artifact: no native MTP speculative head and no resident vision tower weights, even though config.json may still list mtp_* / vision_config fields copied from the source config.

Folder name retains -mtp8 for pipeline naming consistency with other Colibro Qwen converts; that does not imply MTP tensors are present in this upload.

Key features

• INT4 MoE weights for AgentWorld-style agent / environment workloads under Colibro
• Same Colibro shard layout and runtime path as other Qwen MoE converts
• Tokenizer and generation configs carried from the source tree

Files (ballpark)

Pattern	Notes
colibro-00000.safetensors … colibro-00020.safetensors	Weight shards (~16.8 GB combined)
config.json	Model config
tokenizer.json, tokenizer_config.json, vocab.json, merges.txt	Tokenizer
generation_config.json	Sampling defaults
preprocessor_config.json, video_preprocessor_config.json	Present from source; no converted vision weights

How to run (Colibro)

Clone / build Colibro, then point the Qwen runtime at a local copy of this repo. Usage patterns match the Qwen docs (README-QWEN.md, QUICKSTART-QWEN.md); swap the model path and set a distinct --model-id if you serve alongside other checkpoints.

Download

huggingface-cli download Colibro/Qwen-AgentWorld-35B-A3B-INT4 `
  --local-dir models\Qwen-AgentWorld-35B-A3B-colibro-int4-mtp8

Doctor / validate

python c\coli doctor `
  --runtime qwen `
  --model models\Qwen-AgentWorld-35B-A3B-colibro-int4-mtp8 `
  --auto-tier

Web UI + API

python c\coli web `
  --runtime qwen `
  --model models\Qwen-AgentWorld-35B-A3B-colibro-int4-mtp8 `
  --model-id qwen-agentworld-35B-COLIBRO-INT4 `
  --auto-tier `
  --gpu auto `
  --kv-slots 4 `
  --port 8000

CLI one-shot

python c\coli run `
  --runtime qwen `
  --model models\Qwen-AgentWorld-35B-A3B-colibro-int4-mtp8 `
  --auto-tier `
  --gpu auto `
  --ngen 128 `
  "Describe a short agent environment interaction."

Intended use / limitations

• Use with Colibro (coli run / coli web / coli serve / qwen.exe). Loading these shards in stock Transformers AutoModel is unsupported.
• No MTP head and no vision weights in this convert — text generation only under Colibro.
• Quantization trades some quality vs the original weights for smaller storage and Colibro tiered expert placement.

Attribution

Converted from Qwen/Qwen-AgentWorld-35B-A3B (Alibaba Qwen team; base lineage includes Qwen3.5-35B-A3B-Base / AgentWorldBench) using Colibro’s convert_qwen_moe.py pipeline.

Please cite and respect the license of the base model when redistributing or publishing derivatives.