PolyKode-Enterprise-Base

Instructions to use ProCogia/PolyKode-Enterprise-Base with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use ProCogia/PolyKode-Enterprise-Base with Transformers:

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

pipe = pipeline("text-generation", model="ProCogia/PolyKode-Enterprise-Base", trust_remote_code=True)
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoProcessor, AutoModel

processor = AutoProcessor.from_pretrained("ProCogia/PolyKode-Enterprise-Base", trust_remote_code=True)
model = AutoModel.from_pretrained("ProCogia/PolyKode-Enterprise-Base", trust_remote_code=True, device_map="auto")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = processor.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(processor.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use ProCogia/PolyKode-Enterprise-Base with vLLM:

Install from pip and serve model

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

Use Docker

docker model run hf.co/ProCogia/PolyKode-Enterprise-Base

SGLang

How to use ProCogia/PolyKode-Enterprise-Base 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 "ProCogia/PolyKode-Enterprise-Base" \
    --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": "ProCogia/PolyKode-Enterprise-Base",
		"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 "ProCogia/PolyKode-Enterprise-Base" \
        --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": "ProCogia/PolyKode-Enterprise-Base",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use ProCogia/PolyKode-Enterprise-Base with Docker Model Runner:
```
docker model run hf.co/ProCogia/PolyKode-Enterprise-Base
```

PolyKode Enterprise — Base

PolyKode Enterprise Base is ProCogia's base checkpoint for its commercial code-translation product family. This release is the byte-identical, unmodified weights of moonshotai/Kimi-K2.7-Code by Moonshot AI, republished under the same Modified MIT License as the foundation on which ProCogia builds PolyKode Enterprise's fine-tuned adapters, parity harness, and audit trails.

This release is the base checkpoint only. No PolyKode fine-tuning is applied in these weights. It is the substrate we build Enterprise on — not a substitute for it. For the fully-adapted, execution-parity-certified PolyKode Enterprise product with SAS 9.4 / Viya certification, audit trails, and regulated-delivery SLAs, see PolyKode Enterprise or contact outreach@procogia.com.

🌐 procogia.com · ✉️ outreach@procogia.com · 🔒 SOC 2 Type 1 attested · 🇨🇦 Vancouver, BC · North America

Why Kimi K2.7 Code

Kimi K2.7 Code was chosen as the Enterprise base for three reasons:

Coding-agent orientation — Moonshot AI trained K2.7 Code on real-world long-horizon coding tasks, with substantial improvements on end-to-end task completion across complex software-engineering workflows compared to prior Kimi checkpoints.
Efficient inference — token efficiency is meaningfully better than K2.6, with roughly 30% less thinking-token usage on comparable tasks. That matters for regulated deployments where inference cost and latency have to be predictable.
Compatible license — Modified MIT permits commercial redistribution and modification. The only additional condition is a UI attribution requirement at very high scale (100M MAU or $20M/month revenue). See License below.

Community vs Enterprise vs Enterprise-Base

	PolyKode Community	PolyKode Enterprise-Base (this repo)	PolyKode Enterprise
Base weights	Qwen3-32B	Kimi K2.7 Code	Kimi K2.7 Code + proprietary adapters
PolyKode SFT / RLVR training	❌	❌	✅ Included
Execution-parity harness	❌	❌	✅ Full harness
SAS 9.4 / Viya certification	❌	❌	✅ Included
SAS-exact numerical contract	❌	❌	✅ Enforced
Audit trails	❌	❌	✅ Generated per translation
Deployment	Self-serve HF	Self-serve HF	ZeroBoxx on-prem / sovereign cloud / dedicated endpoint
Support	Community	Community	Named engineering + product contact
License	Apache 2.0	Modified MIT	ProCogia Enterprise EULA
Repo	PolyKode-Community	This repo	PolyKode-Enterprise

If you are experimenting with PolyKode workflows or benchmarking, use Community (Qwen3-32B, Apache 2.0) or this Base (Kimi K2.7 Code, Modified MIT). If you need execution-parity certification, audit trails, and SLAs, engage on Enterprise.

Model description

Architecture: Kimi K2 (Mixture of Experts, coding-agent focused)
Base parameters / activation: see upstream Kimi K2.7 Code card for exact MoE specifics
Context length: long-context (see upstream card for current limit)
Precision: compressed-tensors safetensors, 64 shards (~595 GB on disk)
Tokenizer: Kimi tokenizer (upstream)
Task focus: coding-agent workflows, long-horizon software engineering, end-to-end task completion

This repository contains the unmodified upstream weights. All architectural, training, and evaluation details are the responsibility of Moonshot AI; refer to moonshotai/Kimi-K2.7-Code for the authoritative technical description.

Intended use

Good fits:

Evaluating base coding-agent capability before engaging on PolyKode Enterprise
Prototyping PolyKode Enterprise integrations on your own infrastructure
Baselines for SAS / R / Python translation research using PolyKode-style parity harnesses
General Kimi K2.7 Code use cases (agentic coding, long-horizon tasks) with ProCogia branding

Not intended for:

Production SAS → R / Python translation in regulated environments — use PolyKode Enterprise
Any use case requiring execution-parity guarantees, audit trails, or an SLA
Substituting for licensed SAS runtime validation

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "ProCogia/PolyKode-Enterprise-Base"
tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype="auto",
    device_map="auto",
    trust_remote_code=True,
)

For serving at scale we recommend vLLM or SGLang with appropriate tensor-parallel and expert-parallel settings for a ~1T-parameter MoE architecture. Contact ProCogia for a reference deployment for ZeroBoxx.

Attribution

This model is a byte-identical republication of moonshotai/Kimi-K2.7-Code by Moonshot AI, released under the Modified MIT License. Only the repository branding, model card, and metadata differ; no weight modifications have been made in this release.

Original work: moonshotai/Kimi-K2.7-Code
Original copyright: Copyright (c) 2026 Moonshot AI
Original license: Modified MIT (see LICENSE)
Modifications: Repository branding, model card, and metadata identify this artifact as PolyKode Enterprise-Base by ProCogia.

We are grateful to the Moonshot AI team for open-sourcing Kimi K2.7 Code under a permissive commercial license, making this Enterprise-Base distribution possible.

License

Licensed under the Modified MIT License. See LICENSE in this repository for the full text.

Key terms:

✅ Free commercial use, modification, redistribution, and sublicensing are permitted
📝 Copyright and license notice must be preserved in all copies and substantial portions
⚠️ UI attribution at scale: If you use this Software (or any derivative works) in commercial products or services with more than 100 million monthly active users or more than $20 million USD in monthly revenue, you must prominently display "Kimi K2.7 Code" on that product's or service's user interface.
🚫 Provided "as-is" — no warranty; Moonshot AI and ProCogia disclaim liability

The UI attribution requirement inherits to downstream users of this repository. If your deployment plausibly meets the thresholds above, plan for that UI treatment; contact your legal team if unsure.

Citation

If you use PolyKode Enterprise-Base in research, cite both the upstream Kimi work and PolyKode:

@misc{kimi_k27_code,
  title  = {Kimi K2.7 Code},
  author = {Moonshot AI},
  year   = {2026},
  url    = {https://huggingface.co/moonshotai/Kimi-K2.7-Code}
}

@misc{polykode2026enterprisebase,
  title  = {PolyKode Enterprise-Base: coding-agent base checkpoint for regulated code translation},
  author = {ProCogia},
  year   = {2026},
  url    = {https://huggingface.co/ProCogia/PolyKode-Enterprise-Base}
}

🌐 ProCogia on Hugging Face — organization page
🎯 PolyKode Enterprise — commercial product page (parity harness, SAS certification, SLAs)
🔀 PolyKode Community — Apache 2.0 base checkpoint (Qwen3-32B)
🌐 procogia.com
✉️ outreach@procogia.com

_{ProCogia · Founded 2013 · Headquartered in Vancouver, British Columbia · SOC 2 Type 1 attested · This release is byte-identical to moonshotai/Kimi-K2.7-Code. Upstream license is preserved; the UI-attribution threshold clause applies to all downstream users.}

Downloads last month: 127

Safetensors

Model size

1T params

Tensor type

BF16

F32

I32

Model tree for ProCogia/PolyKode-Enterprise-Base

Base model

moonshotai/Kimi-K2.7-Code

Finetuned

(10)

this model