Modelpublic

OsaurusAI/Qwen-AgentWorld-35B-A3B-MXFP8

sourceHugging Faceapache-2.0updated 3mo agoView on Hugging Face

5likes88downloads

Model Card

[image]

Qwen-AgentWorld-35B-A3B · MXFP8

Official OsaurusAI MXFP8 build of Qwen/Qwen-AgentWorld-35B-A3B (Apache-2.0) — a ~35B-total / ~3B-active hybrid MoE (qwen3_5_moe). Near-lossless 8-bit microscaled FP; runs on Apple Silicon via Osaurus / mlx_lm.

—~33 GB bundle (down from ~65 GB bf16).
—MXFP8: microscaled FP8 (group-size 32) across the weights, high-precision embeddings/head. Use this for the highest fidelity; use JANG_4M (~17.6 GB) for the smallest footprint.
—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).

Reasoning behavior. This model shows its work ("Thinking Process: …") by default. The soft /no_think suppression switch is not reliable on this MXFP8 pack — for deterministic reasoning-off, prefer the JANG_4M pack, which honors it.

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

bash

python -m mlx_lm generate --model OsaurusAI/Qwen-AgentWorld-35B-A3B-MXFP8 --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 · MXFP8 (microscaled FP8, group-size 32) · eric@osaurus.ai