Official OsaurusAI mixed-precision (JANG_4M) build of Qwen/Qwen-AgentWorld-35B-A3B (Apache-2.0) — a ~35B-total / ~3B-active hybrid MoE (qwen3_5_moe). Quantized by Osaurus; runs on Apple Silicon via Osaurus / mlx_lm.

~17.6 GB bundle (down from ~65 GB bf16).
JANG_4M: 8-bit affine attention + router, 4-bit affine routed experts, high-precision embeddings/head.
Text-only. The upstream checkpoint ships a vestigial vision_config with no vision-tower weights, so this bundle is correctly stamped modality: text (has_vision: false).

Architecture


Family	`qwen3_5_moe` (hybrid)
Layers	40 — 30 linear-attention (Gated DeltaNet / SSM) + 10 full-attention (1 every 4)
Experts	256 routed (8 active)
Active params	~3 B
Cache	hybrid (recurrent GDN state + KV for full-attn layers)

The Gated-DeltaNet layers (conv1d / A_log / dt_bias / gated in_proj_{qkv,a,b,z}) carry a recurrent state across the full-attention layers — verified coherent with long-context recall in the Osaurus (vMLX-Swift) runtime.

Usage

python -m mlx_lm generate --model OsaurusAI/Qwen-AgentWorld-35B-A3B-JANG_4M --prompt "Explain a hash map in two sentences."

Or load in Osaurus for a local agent loop (tool-calling supported via the qwen tool parser).

Provenance

Quantization: Osaurus · JANG_4M (8-bit attn / 4-bit experts, group-size 32) · eric@osaurus.ai