Qwen-AgentWorld-35B-A3B-bf16
Full-precision (bfloat16) MLX build of
Qwen/Qwen-AgentWorld-35B-A3B.
The original weights load in mlx-lm / oMLX as-is — this repo packages them for
Apple Silicon without any precision loss.
For Apple Silicon. Runs in mlx-lm, oMLX, or any MLX app. Quantized siblings:
oQ4
(≈4.6 bpw, ≈19 GB) and
oQ3.5
(≈3.5 bpw, ≈16 GB).
Notes
- Text-only. The checkpoint declares a
vision_configand MTP heads inconfig.json, but ships no vision ormtp.*weights (693 tensors, 0 vision, 0 MTP) — both are vestigial skeleton inherited from the Qwen3.5 base, andlanguage_model_only: truemakes it a pure LM. Nothing multimodal is present to use. - ≈65 GB on disk; peak memory ≈69 GB generating. Fits in 128 GB unified memory; use a quantized sibling on smaller machines.
Performance
Measured with oMLX (Auto engine) on a Macbook Pro M5 Max 128GB 40 GPU. Single request:
| Context (pp/tg) | TTFT | decode | prefill | peak mem |
|---|---|---|---|---|
| 1024 / 128 | 644 ms | 77 tok/s | 1591 tok/s | 65.6 GB |
| 4096 / 128 | 1.68 s | 76 tok/s | 2434 tok/s | 66.4 GB |
| 8192 / 128 | 2.39 s | 75 tok/s | 3428 tok/s | 66.7 GB |
| 32768 / 128 | 12.0 s | 67 tok/s | 2730 tok/s | 68.7 GB |
For ≈1.8× the single-request decode throughput at ≈⅓ the memory, see the oQ4 sibling.
Usage
mlx_lm.generate --model mlx-community/Qwen-AgentWorld-35B-A3B-bf16 \
--system-prompt "You are a language world model simulating a Linux terminal. Given the user's command, predict the terminal output." \
--prompt $'Action: execute_bash\nCommand: ls -la /home/user/project/' \
--max-tokens 512 --temp 0.6
The model uses thinking mode (<think>...</think>) by default. Recommended
sampling: temperature=0.6, top_p=0.95, top_k=20. See the
base model card for the
seven agent domains and domain-specific system prompts.