WIA-35B-A3B

WIA-35B-A3B is a 35B-total-parameter, approximately 3B-active-parameter MoE world model from the World Intelligence Agent (WIA) project. It predicts action-conditioned observations across Android, MCP, desktop OS, Search, SWE, Terminal, and Web. It is intended for world-model evaluation and as a synthetic environment for agent training and test-time action lookahead; it is not a general-purpose chat assistant.

The model is initialized from a Qwen3.5-35B-A3B backbone and trained with WIA-CPT, observation-only WIA-SFT, and GSPO-based world-model RL (WIA-WM).

Evaluation

Results below are from the WIA technical report. Scores are on a 0--100 scale, evaluated by DeepSeek-V4-Pro; Avg is sample-weighted and MAvg is the macro average over seven domains.

WIA-Bench OOD

The complete OOD comparison from the technical report is reproduced below. Scores use the DeepSeek-V4-Pro judge; Avg is sample-weighted and MAvg is the macro average over seven domains.

Group	Model	Android	MCP	OS	Search	SWE	Terminal	Web	Avg	MAvg
Proprietary	GPT-5.6-luna	48.37	44.51	54.96	40.01	55.88	62.42	53.87	51.91	51.43
Proprietary	GPT-5.6-terra	48.34	43.55	56.14	40.46	57.68	61.79	54.57	52.38	51.79
Proprietary	GPT-5.6-sol	49.39	45.75	59.70	39.99	58.61	63.63	55.00	53.58	53.15
Proprietary	Claude-Sonnet-5	47.76	49.85	63.22	35.26	44.58	46.65	54.24	48.03	48.79
Proprietary	Claude-Opus-4.8	48.98	53.97	63.02	39.10	58.15	57.93	55.26	53.40	53.77
Open-source	GLM-5.3	47.20	50.23	55.58	31.10	32.68	53.88	51.03	44.98	45.96
Open-source	MiMo-v2.5-pro	45.80	46.18	59.78	38.20	49.73	53.43	51.50	49.09	49.23
Open-source	Qwen3.8-27B	41.08	39.40	53.34	34.88	48.47	51.84	47.51	45.46	45.22
Open-source	Qwen3.8-max	47.60	45.32	57.97	36.82	57.62	63.42	55.48	52.40	52.03
Open-source	DeepSeek-V4-Flash	46.74	44.46	58.93	38.40	50.51	55.53	50.77	49.40	49.33
Open-source	DeepSeek-V4-Pro	47.87	43.92	61.93	38.77	51.14	57.77	52.63	50.73	50.57
Qwen3.5	Qwen3.5-4B	39.92	38.38	55.29	30.44	37.26	39.50	45.24	40.46	40.86
Qwen3.5	Qwen3.5-9B	43.00	41.37	58.19	30.99	44.05	43.81	47.42	43.76	44.12
Qwen3.5	Qwen3.5-27B	47.84	42.56	64.29	33.75	48.10	55.09	52.78	49.15	49.20
Qwen3.5	Qwen3.5-35B-A3B	47.09	42.69	56.59	32.03	45.70	49.80	49.55	46.01	46.21
Qwen3.5	Qwen3.5-122B-A10B*	43.92	32.43	41.86	34.93	48.65	55.70	44.15	44.07	43.09
Qwen3.5	Qwen-AgentWorld	47.43	42.63	64.21	45.02	56.31	57.22	53.68	52.87	52.36
WIA-WM	WIA-WM-4B	27.19	44.10	48.40	36.84	60.06	55.87	49.27	46.34	45.96
WIA-WM	WIA-WM-9B	46.82	52.12	62.70	46.06	60.77	59.38	53.43	54.49	54.47
WIA-WM	WIA-WM-27B	50.51	52.53	63.42	46.36	59.92	60.73	54.52	55.45	55.43
WIA-WM	WIA-WM-35B-A3B	48.76	43.72	63.86	51.33	59.19	62.52	54.95	55.72	54.90
WIA-WM	WIA-WM-122B-A10B*	50.59	40.00	64.71	58.18	60.46	64.62	53.59	57.35	56.02

WIA-Bench ID

The complete ID comparison from the technical report is reproduced below.

Group	Model	Android	MCP	OS	Search	SWE	Terminal	Web	Avg	MAvg
Proprietary	GPT-5.6-sol	43.62	61.02	33.08	50.66	53.11	59.96	62.38	52.04	51.98
Proprietary	GPT-5.6-terra	41.57	61.98	33.08	47.32	52.81	59.25	60.94	51.05	50.99
Proprietary	GPT-5.6-luna	41.29	60.03	32.43	45.40	51.29	55.28	58.33	49.11	49.15
Proprietary	Claude-Opus-4.8	41.59	59.08	33.00	63.50	56.46	65.91	66.21	55.76	55.11
Proprietary	Claude-Sonnet-5	40.08	60.28	32.27	58.46	38.72	60.40	63.96	50.75	50.59
Open-source	GLM-5.3	41.20	58.63	33.20	54.74	32.09	63.22	64.52	49.64	49.66
Open-source	MiMo-v2.5-pro	40.93	59.61	32.39	51.34	47.36	54.35	64.16	50.12	50.02
Open-source	Qwen3.8-27B	36.69	55.67	30.75	50.87	43.50	52.67	61.15	47.59	47.33
Open-source	Qwen3.8-max	41.18	58.32	32.39	61.00	51.06	65.09	65.82	54.06	53.55
Open-source	DeepSeek-V4-Flash	40.64	59.23	32.39	56.41	44.87	56.97	64.40	50.90	50.70
Open-source	DeepSeek-V4-Pro	41.54	59.06	32.31	55.72	45.65	55.51	64.51	50.76	50.61
Qwen3.5	Qwen3.5-4B	35.61	54.57	30.96	45.63	33.14	40.22	56.33	42.20	42.35
Qwen3.5	Qwen3.5-9B	36.80	55.09	31.81	38.29	37.16	38.97	60.23	42.39	42.62
Qwen3.5	Qwen3.5-27B	41.24	57.26	32.09	45.92	43.77	57.37	64.72	48.91	48.91
Qwen3.5	Qwen3.5-35B-A3B	37.50	56.75	32.21	39.15	41.76	50.03	61.36	45.47	45.54
Qwen3.5	Qwen-AgentWorld	39.03	56.99	32.49	63.68	45.60	61.33	64.54	52.51	51.95
WIA-WM	WIA-WM-4B	27.12	65.26	30.76	82.59	57.52	75.76	63.24	59.17	57.46
WIA-WM	WIA-WM-9B	41.61	63.89	32.79	84.31	60.87	73.96	68.69	61.99	60.87
WIA-WM	WIA-WM-27B	43.21	67.74	35.28	88.54	66.24	80.85	73.84	66.41	65.10
WIA-WM	WIA-WM-35B-A3B	43.49	67.45	34.39	89.30	64.18	76.36	71.86	65.04	63.86
WIA-WM	WIA-WM-122B-A10B*	41.44	64.33	33.01	86.66	57.74	71.43	66.21	61.16	60.12

*The 122B-A10B configuration differs from the other WIA-WM scales; it is RL-only and uses a different RL data and reward setup.

Stage ablation

OOD-only ablation under the DeepSeek-V4-Pro judge:

Stage	Android	MCP	OS	Search	SWE	Terminal	Web	Avg
Qwen3.5-35B-A3B-Base	35.08	31.83	43.62	27.35	33.70	36.21	37.41	34.93
WIA-CPT-35B-A3B	30.04	39.82	56.37	24.48	32.71	41.39	49.20	38.40
WIA-SFT-35B-A3B	43.69	40.08	52.21	40.13	50.16	54.33	51.59	47.96
WIA-WM-35B-A3B	48.76	43.72	63.86	51.33	59.19	62.52	54.95	55.72

World-model RL adds 7.76 OOD points and 5.51 ID points over WIA-SFT-35B-A3B. The model is strongest on ID Search (89.30) and is competitive with the 27B model despite activating only about 3B parameters per token.

Simulation experiments

The technical report uses WIA-WM-35B-A3B as the synthetic environment, served on 32 GPUs at temperature 0.6 and top-p 0.95. The policy is Qwen3.5-4B, and evaluation uses at most 20 interaction turns.

Simulation SFT and RL

The following results compare the Qwen3.5-4B prompt baseline with synthetic SFT (+SSFT) and synthetic RL (+SRL). Scores are percentages; the BFCL-v3 and tau2 columns are separate benchmark subsets.

Method	BFCL Non-Live	BFCL Live	BFCL Multi-Turn	BFCL Hallu.	BFCL Relev.	tau2 Airline	tau2 Retail	tau2 Telecom
Prompt	36.65	31.46	6.75	42.32	62.50	40.00	33.33	0.00
+SSFT	40.92	37.23	28.00	62.17	81.25	48.00	42.98	16.67
+SRL	46.23	41.15	19.88	63.32	81.25	38.00	65.79	16.67

Relative to the prompt baseline, +SSFT lifts the BFCL-v3 overall by 13.98 points and tau2 by 11.44 points; +SRL lifts them by 14.43 and 15.71 points, respectively. +SRL is stronger on BFCL Non-Live/Live and tau2 Retail, while +SSFT retains an advantage on BFCL Multi-Turn and tau2 Airline.

Test-time inference lookahead

At inference time, the policy samples and de-duplicates three candidate actions. WIA-WM-35B-A3B predicts the next observation for each candidate, and a temperature-0 selector chooses the most promising action before only that action is executed in the real environment. The policy is DeepSeek-V4-Pro; no policy retraining is used.

OOD benchmark	Maximum turns	Accuracy gain	Average successful-trajectory turn change
TerminalBench 2.0	50	+6.74	+2.93
AndroidWorld	20	+4.31	-0.04
BrowseComp-Plus	30	+3.73	-3.39
WebArena	30	+2.69	-0.67

Negative turn changes indicate fewer average turns on successful trajectories; positive values indicate more turns.

Download

pip install -U "transformers>=5.8.0" huggingface_hub torch
hf download darwinagent/WIA-35B-A3B --local-dir ./WIA-35B-A3B

Usage

WIA-35B-A3B predicts an observation after an action. Keep the WIA prompt format and append ###[Observation] before generation.

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "darwinagent/WIA-35B-A3B"
tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True,
)

messages = [
    {"role": "system", "content": "Operate as a language world model and report only the resulting observation."},
    {"role": "user", "content": "Task: inspect the repository.\nCurrent observation: the shell is open.\nAction: shell_exec(command='pwd', cwd='/workspace', timeout_sec=30)\n###[Observation]"},
]
inputs = tokenizer.apply_chat_template(
    messages, tokenize=True, add_generation_prompt=True, return_tensors="pt"
).to(model.device)
with torch.inference_mode():
    output = model.generate(inputs, max_new_tokens=16384, do_sample=False)
print(tokenizer.decode(output[0, inputs.shape[-1]:], skip_special_tokens=True))

Use a GPU configuration appropriate for the model and context length. Practical memory use depends on dtype, context length, batching, and attention implementation.

Citation

Please cite the WIA technical report when using this model.