leonsarmiento/Qwen-AgentWorld-35B-A3B-6bit-XL-mlx

This model was converted to MLX format from Qwen/Qwen-AgentWorld-35B-A3B using BaseQuant_XL 6/8-bit mixed quantization optimized for Apple Silicon.

BaseQuant_XL keeps the most routing-critical layers in full bf16 precision — the MoE router gate, shared expert gate, shared expert, and lm_head — while applying aggressive quantization to the bulk parameters. This preserves routing accuracy and output quality where it matters most.

About the Source Model

Qwen-AgentWorld-35B-A3B is the first language world model — a model that simulates agentic environments via long chain-of-thought reasoning. Instead of executing actions, it predicts the next environment state given an agent's action and interaction history.

Key facts:

• Architecture: Qwen3.5-35B-A3B-Base MoE (35B total / 3B active, 256 experts, 8 active + 1 shared)
• Training: CPT → SFT → RL (GSPO) — environment modeling is the core training objective, not a post-hoc adaptation
• Domains: 7 unified agent interaction domains — MCP (tool calling), Search, Terminal, SWE, Android, Web, OS
• Context: 262,144 tokens
• Output: Predicts next environment observation using <think>...</think> reasoning

Use with mlx

pip install -U mlx-vlm

python -m mlx_vlm.generate --model leonsarmiento/Qwen-AgentWorld-35B-A3B-6bit-XL-mlx --max-tokens 4096 --temperature 0.6 --top-p 0.95 --top-k 20 --prompt "Hello"

BaseQuant_XL Quantization Strategy

Bit Depth	Layers	Rationale
bf16 (unquantized)	mlp.gate (router), shared_expert_gate, lm_head, shared_expert	Routing decisions and shared computation path — errors here are qualitatively different from precision loss
8-bit	embed_tokens, self_attn (full attention), linear_attn (DeltaNet)	Every-token layers with moderate sensitivity — 8-bit is near-lossless
6-bit	switch_mlp (routed experts)	Bulk of parameters, only 8 of 256 experts active per token — natural redundancy tolerates lower precision

Quantization Details

Layer	Bits	Group Size
mlp.gate (router)	bf16	—
shared_expert_gate	bf16	—
lm_head	bf16	—
shared_expert	bf16	—
embed_tokens	8	64
self_attn (full attention)	8	64
linear_attn (DeltaNet)	8	64
switch_mlp (routed experts)	6	64
Default fallback	8	64

• Quantization type: BaseQuant_XL mixed (text-only, vision tower absent)
• Bits per weight: 6.892
• Total size: ~28 GB (6 shards)
• Group size: 64
• Framework: mlx_vlm with strict=False (text-only checkpoint with VLM architecture — same pattern as KAT-Coder-V2.5-Dev)

Recommended Inference Parameters

Parameter	Value
temperature	0.6
top_p	0.95
top_k	20
max_tokens	32,768 (increase for multi-step trajectories)
reasoning_parser	qwen3
enable_thinking	true (default)

Preserve thinking: This is a Qwen3.5 model. In LM Studio, enable preserve_thinking to capture reasoning content separately from output. Set chat_template_kwargs: {"preserve_thinking": true} in API payloads.

Text-only note: Although the config includes vision_config, the checkpoint contains only language model weights (no vision tower). Use with strict=False or --language-model-only in frameworks that check for vision components.

Flight Simulator Demo

AgentWorld-35B-A3B coding a flight simulator from scratch.

MLX Benchmarks (n=50, instruct mode)

Benchmark	AgentWorld 6bit-XL	Qwen3.6-35B 6bit-XL	Agents-A1 6bit-XL	Ornith-35B 6bit-XL	Huihui-Qwen3.6 6bit-XL	Gemma-4 26B 6bit-XL
MMLU	76%	68%	66%	64%	74%	76%
MMLU_PRO	66%	60%	58%	66%	66%	82%
ARC_CHALLENGE	74%	90%	90%	92%	92%	90%
HUMANEVAL	82%	74%	90%	78%	84%	98%
MBPP	78%	80%	70%	78%	78%	82%
Average	75.2%	74.4%	74.8%	75.6%	78.8%	85.6%

AgentWorld leads the Qwen3.6 family on MMLU (+8 over base) and MMLU_PRO (+6), reflecting its deep training on agent interaction domains that overlap with knowledge and reasoning tasks. The ARC_CHALLENGE drop (74% vs 90%) is notable — the world-model fine-tuning trades commonsense reasoning strength for environment simulation capability. Coding benchmarks (HUMANEVAL 82%, MBPP 78%) are strong, consistent with its SWE/Terminal domain coverage.

About XL Quantization

BaseQuant_XL is a data-agnostic (static) quantization method — it allocates precision based on layer role in the architecture, not calibration data. Unlike data-dependent methods (iMatrix, AWQ, GPTQ, oQ, oQ4e) which use calibration datasets that can skew representation toward well-represented domains, XL quantization treats all tokens equally. This makes it suitable for general-purpose use without domain bias.