Instructions to use AlephFunk/Qwen3-8B-Base-del14 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use AlephFunk/Qwen3-8B-Base-del14 with Transformers:

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

pipe = pipeline("text-generation", model="AlephFunk/Qwen3-8B-Base-del14")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("AlephFunk/Qwen3-8B-Base-del14")
model = AutoModelForCausalLM.from_pretrained("AlephFunk/Qwen3-8B-Base-del14", device_map="auto")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use AlephFunk/Qwen3-8B-Base-del14 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "AlephFunk/Qwen3-8B-Base-del14"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "AlephFunk/Qwen3-8B-Base-del14",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/AlephFunk/Qwen3-8B-Base-del14

SGLang

How to use AlephFunk/Qwen3-8B-Base-del14 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 "AlephFunk/Qwen3-8B-Base-del14" \
    --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": "AlephFunk/Qwen3-8B-Base-del14",
		"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 "AlephFunk/Qwen3-8B-Base-del14" \
        --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": "AlephFunk/Qwen3-8B-Base-del14",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use AlephFunk/Qwen3-8B-Base-del14 with Docker Model Runner:
```
docker model run hf.co/AlephFunk/Qwen3-8B-Base-del14
```

Qwen3-8B-Base-del14

Qwen3-8B-Base with decoder layer 14 deleted (36 -> 35 layers). Tier-1 'safest' cut from the laco-lab study.

This is a research artifact, not a recovered model. It is a base (pretrained) Qwen3 model with decoder layers removed and no healing / fine-tuning of any kind. Its language-modelling loss and especially its state-tracking and multi-step-reasoning ability are degraded relative to Qwen/Qwen3-8B-Base; it is intended as a starting point for recovery experiments (e.g. recurrent replacement of the removed span).

Parent: Qwen/Qwen3-8B-Base (a base checkpoint; no instruct/post-trained weights were used).
Deleted original layers: [14].
Surviving layers are bit-identical to the parent's (renumbered), except where a collapsed group is listed (then: last layer's attention/MLP weights, base layer's norms).
Standard Qwen3 architecture: AutoModelForCausalLM.from_pretrained("AlephFunk/Qwen3-8B-Base-del14"). config.json has the corrected num_hidden_layers and layer_types.
Exact recipe: laco_recipe.json (ordered merge/delete ops); provenance with shard SHA-256 hashes: provenance.json.

Measured effect (paired against the parent, identical evaluation items)

Parent vs this model, precision used for both: NF4 (bitsandbytes, lm_head also NF4). Evaluation suite digest 3865a9dda944, sweep-size evaluation (PIQA/ARC: 400 items each, zero-shot log-likelihood; synthetic probes: 160 items each, few-shot, seeded).

metric	parent	this model
wikitext-2 test perplexity (1024-token windows)	7.919	8.358
pile-10k sample perplexity (512 tokens)	7.385	7.788
PIQA acc_norm	0.772	0.775
ARC-Easy acc_norm	0.777	0.805
ARC-Challenge acc_norm	0.532	0.540
swap-tracking (chance 0.25)	0.600	0.519
location-tracking (chance 0.20)	0.956	0.900
multi-step variable arithmetic, ranked (chance 0.20)	0.863	0.831
multi-step greedy-generation exact match (n=80)	0.738	0.738

Paired change in wikitext loss: +0.054 +/- 0.008 nats/token (95% CI). Probe accuracies have 95% CIs of roughly +/-0.04 (PIQA/ARC) to +/-0.12 (synthetic probes); differences smaller than that are noise. Evaluation was run through an in-memory layer view of the parent; on the 8B recipe this was verified to be bit-identical to loading the written checkpoint.

Caveats

No healing; expect degraded fluency and reasoning. Do not use as a drop-in replacement for the parent.
Evaluated only on the tasks above (no MMLU, HellaSwag, WinoGrande, safety or instruction-following evaluation).
Weights inherit the parent's Apache-2.0 license; no additional training data was used.

Provenance and code

Produced by the laco-lab subproject (CLI lacolab, 81 tests, audit of Mercity's LaCo-8B, full report and experiment manifest) in the jev-qwen repository of MoralityLabAI (branch add/laco-lab-layer-compression). Rebuild: lacolab compress --model Qwen/Qwen3-8B-Base --recipe runs/c8/s1.json --out <dir>.

Downloads last month: 23

Safetensors

Model size

8B params

Tensor type

BF16

Model tree for AlephFunk/Qwen3-8B-Base-del14

Base model

Qwen/Qwen3-8B-Base

Finetuned

(634)

this model