Instructions to use msingiai/akilicode-14b-research-preview with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use msingiai/akilicode-14b-research-preview with Transformers:

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

pipe = pipeline("text-generation", model="msingiai/akilicode-14b-research-preview")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("msingiai/akilicode-14b-research-preview")
model = AutoModelForCausalLM.from_pretrained("msingiai/akilicode-14b-research-preview")
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=40)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps

vLLM

How to use msingiai/akilicode-14b-research-preview with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "msingiai/akilicode-14b-research-preview"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "msingiai/akilicode-14b-research-preview",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/msingiai/akilicode-14b-research-preview

SGLang

How to use msingiai/akilicode-14b-research-preview 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 "msingiai/akilicode-14b-research-preview" \
    --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": "msingiai/akilicode-14b-research-preview",
		"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 "msingiai/akilicode-14b-research-preview" \
        --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": "msingiai/akilicode-14b-research-preview",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use msingiai/akilicode-14b-research-preview with Docker Model Runner:
```
docker model run hf.co/msingiai/akilicode-14b-research-preview
```

AkiliCode-14B Research Preview

AkiliCode-14B is an early coding model from MsingiAI released as a research preview.

This release is published as:

msingiai/akilicode-14b-research-preview

Release artifact directory:

/outputs/akilicode-14b-research-preview

Source checkpoint:

/outputs/akili-code-stage3-ckpt100-s50-r4/checkpoint-20

It is a post-trained Qwen2.5-Coder-14B model optimized for structured reasoning and repair-oriented coding tasks. In MsingiAI's current internal evaluation stack, it shows promising function-level coding performance, but it is not yet strong on hidden-test competitive-programming benchmarks.

Research Preview Status

This model is being released for:

research
evaluation
failure analysis
downstream experimentation

This model is not being released as a state-of-the-art coding model or as a strong LiveCodeBench model.

Key Metrics

Promoted checkpoint results:

Benchmark	Result
HumanEval+	62.80
MBPP+	65.61
BigCodeBench-Instruct	45.09
CRUXEval-O	49.75
LiveCodeBench v6 official	11.37

Important Caveat on LiveCodeBench

The main remaining weakness is hidden-test algorithmic correctness, not output parsing.

On the official-style LiveCodeBench v6 run:

n = 1055
pass@1 = 11.37
private tests used for all 1055 problems
extraction_success_rate = 100.0
syntax_valid_rate = 98.58

Failure breakdown:

wrong_answer = 838
timeout = 61
runtime_error = 21
syntax_error = 15
extraction_failed = 0

This means the model is usually producing executable outputs, but it still struggles on hidden-test algorithmic generalization, especially on medium and hard competition-style problems.

Intended Output Format

AkiliCode-14B was trained to respond with two XML blocks in order:

<reasoning>...</reasoning>
<code>...</code>

The reasoning block is expected to contain these headings:

PLAN:
TRACE:
EDGE CASES:
COMPLEXITY:

The code block is expected to contain only executable Python.

If your downstream stack only wants runnable code, extract the contents of <code>...</code> before execution.

Intended Uses

Recommended uses:

structured code generation
code-repair experiments
benchmark research
reasoning-format experiments
evaluation harness development

Less suitable uses right now:

competition-style hidden-test programming
production-critical autonomous coding
benchmark marketing claims about frontier coding performance

How to Use

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

model_id = "msingiai/akilicode-14b-research-preview"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

prompt = """<|im_start|>system
You are Akili Code, an expert programming assistant built by MsingiAI.
Always respond with <reasoning> followed by <code>.
<|im_end|>
<|im_start|>user
Write a Python function that returns the longest palindromic substring of a string.
<|im_end|>
<|im_start|>assistant
"""

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=1024, do_sample=False)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))

Training Summary

AkiliCode-14B uses a three-stage post-training recipe:

supervised fine-tuning
white-box RL for first-pass execution accuracy
repair-focused MURPHY / P-GRPO continuation

The promoted checkpoint was selected because it improved the strongest reliable function-level metrics relative to the Stage 2 golden checkpoint:

HumanEval+: 60.98 -> 62.80
MBPP+: 65.08 -> 65.61

It regressed slightly on BigCodeBench-Instruct:

45.61 -> 45.09

Limitations

weak performance on hidden-test competitive programming
current benchmark profile is much stronger on short function-synthesis tasks than contest-style algorithmic tasks
model behavior depends on downstream handling of the XML output format
not validated for safety-critical or production-critical use

Contact

For research or partnership inquiries:

korir@msingiai.com

Downloads last month: 11

Safetensors

Model size

15B params

Tensor type

BF16

Model tree for msingiai/akilicode-14b-research-preview

Base model

Qwen/Qwen2.5-14B

Finetuned

Qwen/Qwen2.5-Coder-14B

Finetuned

(16)

this model