Qwen-AgentWorld-35B-A3B-MLX-2bit

Qwen-AgentWorld-35B-A3B-MLX-2bit

Summary

MLX 2bit (affine, 2-bit, group size 64) quantization of Qwen/Qwen-AgentWorld-35B-A3B at upstream revision 60d2b0434a53d2e62a7c00a489586815d94ebffb, produced directly from the BF16 safetensors with pipelines.mlx_direct_quantize (majek repo). ~10.1 GiB on disk.

Runtime status (verified 2026-07-04): this pack currently loads under no MLX runtime. mlx-lm 0.31.3's qwen3_5_moe sanitize converts fused expert tensors only in their unquantized, unsuffixed form; this pack's pre-quantized experts.gate_up_proj.{weight,scales,biases} tensors are left unconverted and fail strict loading. Treat this repo as a weights archive until a republish in a loadable layout.

Runtime status

Verified 2026-07-04 with pipelines/lane_a_runtime_verify.py (majek repo); per-runtime outputs and full error traces in .sisyphus/evidence/lane-a-runtime-verify/. Load-tested sibling: 2bit; all eight variants share the same tensor structure.

Republish attempt (2026-07-04): a loadable-layout repack (mlx_lm convert, the route that shipped this repo's 3bit–NVFP4 siblings) was built and smoke-tested but withheld — 2-bit generation degenerates into repetition (.sisyphus/evidence/lane-a-republish/qwen-agentworld-35b-a3b/smoke-2bit.json, majek repo). The siblings were republished in runtime-loadable layout; this repo intentionally remains the original archive pack.

Why this variant

Affine 2-bit, group size 64 — the smallest tier. Highest quality risk of the family (unverified; see the runtime caveat).

Reproduce

python -m pipelines.mlx_direct_quantize --model qwen-agentworld-35b-a3b --base-dir /tmp/mlx-direct-release/qwen-agentworld-35b-a3b/base --out-dir /tmp/mlx-direct-release/qwen-agentworld-35b-a3b/2bit --bits 2 --mode affine --group-size 64

Family

All MLX variants of this model ship together:

• majentik/Qwen-AgentWorld-35B-A3B-MLX-2bit (this repo)
• majentik/Qwen-AgentWorld-35B-A3B-MLX-3bit
• majentik/Qwen-AgentWorld-35B-A3B-MLX-4bit
• 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=2, mode=affine, group_size=64 (text tower only)
• Toolchain: mlx 0.31.2, huggingface_hub 0.36.2
• 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.