Instructions to use fwerkor/CID-v1-0.4B-1018 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use fwerkor/CID-v1-0.4B-1018 with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="fwerkor/CID-v1-0.4B-1018", trust_remote_code=True)
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoModel
model = AutoModel.from_pretrained("fwerkor/CID-v1-0.4B-1018", trust_remote_code=True, device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use fwerkor/CID-v1-0.4B-1018 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "fwerkor/CID-v1-0.4B-1018"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "fwerkor/CID-v1-0.4B-1018",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/fwerkor/CID-v1-0.4B-1018

SGLang

How to use fwerkor/CID-v1-0.4B-1018 with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "fwerkor/CID-v1-0.4B-1018" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "fwerkor/CID-v1-0.4B-1018",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "fwerkor/CID-v1-0.4B-1018" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "fwerkor/CID-v1-0.4B-1018",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use fwerkor/CID-v1-0.4B-1018 with Docker Model Runner:
```
docker model run hf.co/fwerkor/CID-v1-0.4B-1018
```

CID-v1-0.4B-1018

October 2026 retraining release of Continuous Interaction Diffusion (CID).

This is a diffusion-native, tool-augmented reasoning research model, not a standalone autoregressive chat model. The runtime and its custom neural semantics are required.

This release includes the final Stage B checkpoint, all three Stage A epoch-end standalone inference checkpoints, sanitized training logs, and complete paired evaluations of the new and previously published 0.4B models on eight public datasets (44,376 cases each). Under the documented matched evaluation protocol, the new checkpoint improves aggregate Task Accuracy and Answer Discovery Rate, though GSM8K and MATH do not improve. The Task Accuracy numbers are conservative automatic scores with flagged semantic cases still pending adjudication; Answer Discovery Rate is a separate CID-specific mechanism metric, not official benchmark accuracy.

CID 512-example validation (training-time)

Training-time detached free rollout on held-out neural-contract validation; this is not an end-to-end public task benchmark. Datasets and training algorithms may differ from the original CID-v1-0.4B release, so percentages alone are not a controlled between-version comparison.

Checkpoint	Total optimizer steps	Display exact	Need F1	Convergence	Lifecycle	Source
A E1	31,288	6.03%	30.53%	84.48%	37.15%	85.98%
A E2	62,386	12.67%	39.04%	87.90%	85.07%	87.32%
A E3	93,426	15.72%	31.85%	91.11%	91.62%	88.27%
B E1	96,809	18.51%	30.85%	97.54%	98.27%	93.18%

The Stage B checkpoint is the default loaded model. Stage A milestones are frozen-backbone checkpoints reconstructed with the unmodified LFM2.5-Encoder-350M-Diffusion base, the respective CID adapter weights, and the exact frozen semantic snapshot. All were loaded successfully after conversion and checked for non-finite tensor values.

Files

model.safetensors: complete single-file model weights, including CID adapter, Stage B adapted backbone, and semantic embedding.
config.json, cid_config.json, configuration_cid.py, modeling_cid.py, tokenizer files: Hugging Face integration and reproducibility metadata.
checkpoints/stage-a-epoch-{1,2,3}/: standalone inference checkpoints, each with model.safetensors and matching configuration/tokenizer; optimizer states are intentionally excluded.
logs/stage-a/, logs/stage-b/: sanitized numeric training and neural validation logs.
logs/run_config.json, logs/validation_summary.json: exact hyperparameters and summary metrics.

Each complete checkpoint has 212 tensors; model.safetensors is 2,000,402,004 bytes. The model is based on LiquidAI/LFM2.5-Encoder-350M-Diffusion and inherits its LFM license.

Training

Code commit: 3ff8fff67f069b0282645da3ca83eda9ab8a6ee7.
Dataset: fwerkor/CID-Dataset, release/training-trajectories.jsonl; training SHA-256 07662203cc23f5ee628623090ad029740e51b3d6efb13466a6dcad23a2a3b143.
Validation: 512 examples, 3,449 transitions; validation SHA-256 726d864a198302133f9e6005cc14764347a3132f48d5675723b5643d52b754dc.
Hardware: 4 NVIDIA GPUs, BF16 compute.
Stage A: 3 epochs, frozen backbone, CID learning rate 1e-4, effective global batch 96, rollout horizon 3, teacher forcing epoch 1 then rollout ramp 2.
Stage B: 1 epoch, joint backbone+CID training, CID learning rate peak 2e-5, backbone peak 5e-6, WSD-linear with 1% warmup and 10% decay, effective global batch 32, rollout horizon 3, final optimizer step 96,809.
Neural contract v6; semantic head v2; order-aware-v2 pooling. Inference requires a matching CID runtime.

Loading

Install CID from the source repository at the pinned commit or later. For example:

from transformers import AutoModel, AutoTokenizer
model_id = "fwerkor/CID-v1-0.4B-1018"
tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModel.from_pretrained(model_id, trust_remote_code=True)

The model requires the CID runtime to execute cognitive state updates and asynchronous tool scheduling. Loading it does not provide conventional generate() compatibility.

Public task benchmarks: Task Accuracy and Answer Discovery Rate

Full paired benchmarks completed for both releases: all eight datasets, 44,376 examples and 128/128 completed shards per checkpoint (88,752 generated responses in total). No failed shards. Both models were evaluated with the same public dataset SHA-256 values, FP16 Tesla T10 inference, seed 0, --max-steps 64, and 8 denoising steps.

CID reports two complementary metrics, not a replacement for canonical benchmark accuracy:

Task Accuracy (TA): conservative automated detection of an unambiguous correct final answer with known equivalent responses; a correct fact in Known: contradicted by a wrong final answer does not qualify. Unrecognized semantic paraphrases require blinded adjudication.
Answer Discovery Rate (ADR): detection of a correct answer or alias anywhere in generated CID Display, including Known:, even when a later final answer is wrong. ADR measures intermediate answer discovery, not successful user-facing task completion.

Full definitions and regression tests: dual-metric scorer and methodology.

Side-by-side matched model comparison

Dataset	N	Original TA	1018 TA	Original ADR	1018 ADR
ARC-Challenge	1,172	22.53%	30.12%	23.12%	30.55%
MMLU	14,042	15.66%	20.29%	16.25%	21.57%
GSM8K	1,319	6.44%	5.46%	6.44%	5.46%
MATH	4,999	6.90%	6.66%	6.90%	6.72%
MBPP	500	0.00%	0.00%	0.00%	0.00%
HotpotQA	7,405	0.74%	1.04%	0.99%	6.93%
2WikiMultiHopQA	12,576	0.76%	1.48%	1.53%	19.35%
MuSiQue	2,363	0.13%	0.34%	0.25%	0.97%

Micro across 44,376 cases: old TA 6.87% (3,047/44,376), new TA 8.74% (3,878/44,376); old ADR 7.34% (3,255/44,376), new ADR 15.24% (6,764/44,376). Equal-weight macro average across 8 datasets: old TA 6.65%, new TA 8.17%; old ADR 6.94%, new ADR 11.44%. These pooled diagnostics are not official leaderboard scores; the micro average weights large datasets more heavily.

Standard benchmark diagnostics (same-code rescoring)

The following use the benchmark-specific conventional scoring contract for reference, not CID ADR. Historical scores in the original release card were obtained with different evaluation code and should not be conflated with this matched-protocol table.

Dataset	Canonical metric	Original	1018
ARC-Challenge	Strict choice accuracy	21.59%	29.35%
MMLU	Strict choice accuracy	14.94%	18.17%
GSM8K	Numeric exact	6.44%	5.46%
MATH	Normalized math exact	6.90%	6.66%
MBPP	Execution pass@1	0.00%	0.00%
HotpotQA	EM / Token F1	0.74% / 1.58%	0.69% / 6.56%
2WikiMultiHopQA	EM / Token F1	0.76% / 1.59%	1.02% / 13.21%
MuSiQue	EM / Token F1	0.13% / 1.15%	0.34% / 5.80%

Complete result files

The deterministic TA scorer flagged 269 original and 6,020 new-model cases for possible semantic review; flags are not verified correct answers. These headline numbers are the reproducible conservative pre-adjudication scores and must not be described as final human/LLM-confirmed semantic accuracy. Raw per-example trajectories and ambiguity audit files are retained on the evaluation server, not redistributed with benchmark test inputs.

Provenance note: The original baseline finished after a docs/scoring-only Git commit had changed the recorded repository hash for some 2Wiki and MMLU shards; benchmark inference/runtime implementation was unchanged. New and old data digests and runtime configuration were independently cross-checked.

Paper and citation

This checkpoint accompanies the CID paper: Continuous Interaction Diffusion: A Diffusion-Native Architecture for Asynchronous Tool-Augmented Reasoning (arXiv:2608.10438, DOI).

If you use this model or the CID runtime, please cite:

@article{cao2026continuous,
  title   = {Continuous Interaction Diffusion: A Diffusion-Native Architecture for Asynchronous Tool-Augmented Reasoning},
  author  = {Cao, Yuhang and Mu, Yanzhou and Fang, Chunrong and Chen, Zhenyu},
  journal = {arXiv preprint arXiv:2608.10438},
  year    = {2026},
  doi     = {10.48550/arXiv.2608.10438},
  url     = {https://arxiv.org/abs/2608.10438}
}