Qwen-AgentWorld-35B-A3B-MLX-4bit

Qwen-AgentWorld-35B-A3B-MLX-4bit

Summary

MLX 4bit (affine, 4-bit, group size 64) quantization of Qwen/Qwen-AgentWorld-35B-A3B at upstream revision 60d2b0434a53d2e62a7c00a489586815d94ebffb, converted from the BF16 safetensors with mlx_lm convert (mlx-lm 0.31.3). ~18.2 GiB on disk.

Runtime status (verified 2026-07-04): loads and generates under mlx-lm 0.31.3 — chat and code probes passed on this exact payload. Republished 2026-07-04 in the mlx-lm layout, replacing the 2026-07-02 pack, which no MLX runtime could load. Output quality beyond the smoke probes is not verified.

Runtime status

Smoke-gated on Apple Silicon before publish (2026-07-04): this exact payload was loaded and probed per variant (pipelines/vlm_pack_smoke.py / pipelines/lm_pack_smoke.py, majek repo); verdicts in .sisyphus/evidence/lane-a-republish/ and in PROVENANCE.md.

from mlx_lm import load, generate

model, tokenizer = load("majentik/Qwen-AgentWorld-35B-A3B-MLX-4bit")
prompt = tokenizer.apply_chat_template(
    [{"role": "user", "content": "Explain KV caching in one sentence."}],
    add_generation_prompt=True, tokenize=False,
)
print(generate(model, tokenizer, prompt=prompt, max_tokens=128))

Why this variant

Affine 4-bit, group size 64 — the standard MLX deployment tier and the usual quality/size sweet spot.

Reproduce

# base = snapshot of Qwen/Qwen-AgentWorld-35B-A3B @ 60d2b0434a53d2e62a7c00a489586815d94ebffb
python -m mlx_lm convert --hf-path /tmp/lane-a-republish/qwen-agentworld-35b-a3b/base --mlx-path /tmp/lane-a-republish/qwen-agentworld-35b-a3b/4bit -q --q-bits 4 --q-group-size 64 --q-mode affine

Family

All MLX variants of this model ship together:

• majentik/Qwen-AgentWorld-35B-A3B-MLX-2bit
• majentik/Qwen-AgentWorld-35B-A3B-MLX-3bit
• majentik/Qwen-AgentWorld-35B-A3B-MLX-4bit (this repo)
• majentik/Qwen-AgentWorld-35B-A3B-MLX-5bit
• majentik/Qwen-AgentWorld-35B-A3B-MLX-6bit
• majentik/Qwen-AgentWorld-35B-A3B-MLX-8bit
• majentik/Qwen-AgentWorld-35B-A3B-MLX-MXFP4
• majentik/Qwen-AgentWorld-35B-A3B-MLX-NVFP4

Provenance

• Upstream: Qwen/Qwen-AgentWorld-35B-A3B @ 60d2b0434a53d2e62a7c00a489586815d94ebffb
• Quantization: bits=4, mode=affine, group_size=64 (language tower only)
• Toolchain: mlx 0.31.2, mlx-lm 0.31.3, mlx-vlm 0.6.3, transformers 5.12.0, huggingface_hub 1.22.0
• Republished 2026-07-04 in a runtime-loadable layout; supersedes the 2026-07-02 pipelines.mlx_direct_quantize pack. Full details in PROVENANCE.md in this repo.

License + attribution

Quantized by majentik from Qwen/Qwen-AgentWorld-35B-A3B. All rights in the original model remain with its authors.

The upstream model is released under the Apache License 2.0 — see the upstream LICENSE file.