6 min readparanine
Which open-source LLMs you can call from Pakistan today
The open-weight LLMs live on one API from Pakistan in October 2026: chat, OCR, embeddings and decision models, which run in Pakistan, and their rupee prices.
As of 21 October 2026, a developer in Pakistan can call ten open-weight model routes through one OpenAI-compatible API on paranine, five of them served on hardware inside Pakistan, all priced per million tokens in rupees. This post lists them by what they are for.
Catalogues change, so treat this as a dated snapshot: the models page is the current source, and every route ID below is exactly what goes in the model field of a request.
Chat and reasoning
Kimi K3 is the heavyweight for hard reasoning in maths, science and code and for long-horizon agent work. GLM-5.3 is strong on coding, tool use and sustained execution, with GLM-5.3 Fast as a faster, rate-limited preview of the same model. gpt-oss 120B is an Apache 2.0 model for agentic and reasoning-heavy work, and the one to start with if you want a route served in Pakistan. DeepSeek V4.1 Flash is for fast, cost-effective general completions.
| Model | Route ID | Pool |
|---|---|---|
| gpt-oss 120B (OpenAI) | paranine/gpt-oss-120b(Global) | Global |
| gpt-oss 120B (OpenAI) | paranine/gpt-oss-120b(high) | PK NPU |
| Qwen3.8 27B (Qwen) | paranine/qwen-3.8-27b | PK NPU |
| Kimi K3 (Moonshot AI) | paranine/Kimi-K3(Global) | Global |
| GLM-5.3 (Z.ai) | paranine/GLM-5.3(Global) | Global |
| GLM-5.3 Fast (Z.ai) | paranine/GLM-5.3-Fast(Global) | Global |
| DeepSeek V4.1 Flash | paranine/DeepSeek-V4.1-Flash(Global) | Global |
The full catalogue, with every rate per million tokens in rupees
Documents, search and decisions
Three specialised routes run on the PK NPU pool, inside Pakistan.
- DeepSeek-OCR 2 (paranine/DeepSeek-OCR2): send a scan or photo, get structured Markdown back with headings, tables and line items
- Qwen3 Embedding 8B (paranine/qwen3-embedding): production-grade embeddings for RAG and search
- Laya Decision (paranine/laya-decision): ask yes/no, multiple-choice and rating questions about a piece of text and get a probability for each option
What the rupee prices look like
On 21 October 2026, gpt-oss 120B cost Rs 30.08 per million input tokens and Rs 180.47 per million output tokens on either pool, and Qwen3.8 27B cost Rs 80.00 and Rs 650.00. A request costs its prompt tokens at the input rate plus its completion tokens at the output rate, settled against a prepaid rupee balance.
Calling any of them
Every route uses the same endpoint, key and request shape; only the model field changes.
from openai import OpenAI
client = OpenAI(base_url="https://sovereign.paranine.com/v1", api_key="sk_p9_...")
reply = client.chat.completions.create(
model="paranine/gpt-oss-120b(high)",
messages=[{"role": "user", "content": "Summarise this in one line."}],
)Choosing between them
If the data must stay in Pakistan, start from the PK NPU routes and scope the key to them. If it need not, pick by task: the heaviest reasoning model you can afford for hard problems, a Flash or Fast model for volume. Keep the model ID in configuration, so changing it later is not a deploy.
What a sovereign AI gateway is, and how to tell a real one from a badge
Common questions
- Can I use open-source LLMs in Pakistan without hosting them myself?
- Yes. paranine serves open-weight models such as gpt-oss 120B, Kimi K3, GLM-5.3, DeepSeek V4.1 Flash and Qwen3.8 27B through one OpenAI-compatible API, billed per million tokens in rupees. You call a route ID; there is no GPU to rent or model to deploy.
- Which open-source models run inside Pakistan?
- On paranine, the PK NPU-native routes run on Huawei Ascend NPUs inside Pakistan: gpt-oss 120B, Qwen3.8 27B, DeepSeek-OCR 2, Laya Decision and Qwen3 Embedding 8B, as of 21 October 2026. The other live routes are global pass-through and are marked as such in the catalogue.
- What does an open-source LLM cost per million tokens in rupees?
- It depends on the model. On 21 October 2026 gpt-oss 120B was Rs 30.08 per million input tokens and Rs 180.47 per million output tokens on paranine, and Qwen3.8 27B was Rs 80.00 and Rs 650.00. The catalogue at paranine.com/models is the current source for every rate.