CIFUND
Quantitative Finance  ·  AI Inference

Redefining the frontier of quantitative finance and artificial intelligence

Inference

Open-weight language models, served from Brazil

We run open-weight large language models on a dedicated NVIDIA Blackwell GPU (RTX 5090) in Goiânia, Brazil, behind an OpenAI-compatible API built for routers such as OpenRouter. The model is served from NVIDIA's NVFP4 checkpoint (4-bit expert weights, FP8 attention), so it fits a single 32 GB GPU.

Inference location
Goiânia, Brazil (BR)
Hardware
One NVIDIA RTX 5090 (32 GB)
Serving stack
vLLM: NVFP4 weights on Marlin kernels with FP8 attention, speculative decoding, automatic prefix caching
Data handling
Zero data retention: prompts and completions are never logged, stored or used for training. Request metadata is kept 30 days; access records 6 months, as Brazilian law requires.
Model

Qwen3.6 35B A3B, in NVFP4

A 35-billion-parameter hybrid-attention mixture-of-experts model with 3 billion parameters active per token, served from NVIDIA's mixed-precision NVFP4 checkpoint (4-bit expert weights, FP8 attention). The whole model fits one 32 GB GPU. Prices in USD per million tokens. Cached input applies to prompt prefixes already in our cache.

Model Quantization Checkpoint Weights Context Max output Input Cached input Output Status
Qwen3.6 35B A3B
qwen/qwen3.6-35b-a3b
 NVFP4nvidia/Qwen3.6-35B-A3B-NVFP4 23.4 GB262K32K $0.0475$0.0237$0.665 Live

Supports tool calling, structured outputs and reasoning. The machine-readable listing, in OpenRouter's provider document format, is published at api.cifund.xyz/v1/models.

API

OpenAI-compatible endpoints

Base URL   https://api.cifund.xyz/v1
Models     GET  /v1/models
Chat       POST /v1/chat/completions
Text       POST /v1/completions
Auth       Authorization: Bearer <key>

Generation endpoints are live for routing partners (API key required). The models document is public.

  • Streaming: tokens are streamed by server-sent events as they are generated, with keep-alive comments during long prompt processing.
  • Usage: every response, streamed or not, reports prompt, cached and completion tokens.
  • Capacity: at capacity we answer 429 immediately instead of queueing, so routers can retry elsewhere.
  • Features: tool calling, structured outputs (JSON schema) and reasoning controls.
Company

Creative Insight Quant Trading Ltd

Legal entity
Creative Insight Quant Trading Ltd, an International Business Company
Registration
Republic of Seychelles, IBC No. 251741
Operations
GPU inference in Goiânia, Brazil
Contact
will@cifund.xyz