Open-weight models, tuned for agentic workloads: up to 1.25M tokens of context, a warm KV cache across turns, and cached input from $0.006 per 1M tokens.
Served from E2E Networks' high-performance GPU infrastructure.
GLM 5.3 Flash
Z.ai
fp8
GLM-5.3-Flash is a native multimodal model from Z.ai. It is suited for efficient coding and long-horizon agent tasks. Its hybrid sparse and linear attention architecture maintains accurate long-context behavior while...
DeepSeek V4.1 Flash is a sparse mixture-of-experts model from DeepSeek, and the first built on the company's Causal Encoder-Decoder (CED) architecture. It activates 8B parameters on input and 16B on...
Kimi K3 is a 2.8T parameter open-weight multimodal reasoning model from Moonshot AI. It is suited for complex coding, knowledge work, and long-horizon agentic workflows, and is particularly strong at...
Qwen3.8 27B is an open-weight dense vision-language model from Qwen. It is suited for coding, professional workflows, research, multimodal interaction, and long-running agent tasks, with flexible thinking that can be...
Kimi K2.6 is Moonshot AI's next-generation multimodal model, designed for long-horizon coding, coding-driven UI/UX generation, and multi-agent orchestration. It handles complex end-to-end coding tasks across Python, Rust, and Go, and...
Gemma 4 31B Instruct is Google DeepMind's 30.7B dense multimodal model supporting text and image input with text output. Features a 256K token context window, configurable thinking/reasoning mode, native function...
Gemma 4 26B A4B IT is an instruction-tuned Mixture-of-Experts (MoE) model from Google DeepMind. Despite 25.2B total parameters, only 3.8B activate per token during inference — delivering near-31B quality at...
Qwen3.5-9B is a multimodal foundation model from the Qwen3.5 family, designed to deliver strong reasoning, coding, and visual understanding in an efficient 9B-parameter architecture. It uses a unified vision-language design...
Video generation from a prompt and reference media. Guide it with up to 9 images, 3 videos and 3 audio clips, and get back an MP4 through an asynchronous jobs API.
Prices last updated 11 November 2026. Cached input is the price for prompt tokens served from the KV cache; see session affinity & caching.
Built for the workloads that need it most
BlazeInfer is tuned for inference-heavy products where latency and cost per token compound fast.
Coding agents
Low-latency completions and tool calls, with session affinity that keeps multi-turn agent loops on a warm cache instead of starting cold every turn.
Document generation & OCR
Bulk extraction and generation jobs get routed to whichever pod has headroom right now — no hand-tuned load balancer required.
Image & video generation
Bursty, compute-heavy generation workloads spread across our fleet automatically, so you never have to provision GPUs yourself.
The engine underneath
A full inference stack tuned end-to-end — custom kernels, cache, and scheduling — with intelligent routing on top.
Custom inference kernels
Hand-tuned CUDA kernels and a highly optimized serving engine squeeze more tokens per second out of every GPU, at lower latency than stock runtimes.
Unlimited KV cache
Context doesn't get evicted the moment memory tightens. Long conversations and large documents stay warm, so repeat and multi-turn work stays fast and cheap.
Throughput-tuned serving
Continuous batching, scheduling, and memory management tuned specifically for high-concurrency inference — not a generic model server running defaults.
Intelligent routing
Every request lands on the pod with the most headroom and the warmest cache — automatically, across every model we serve.
Case study · Social media
Full-precision Hinglish conversations at 1,000+ RPM
A social media app needed high-quality, multi-turn conversations in Hinglish — without the quality loss its users noticed on quantised models.
BF16
Full-precision Gemma 4 31B
~10K
Tokens per prompt, on average
1,000+
Requests per minute at peak
< 5 s
P95 end-to-end latency
The requirements
Hinglish (Hindi–English code-mixed) content
High-quality Gemma 4 31B generations — quantised versions from other providers led to users dropping off the app
Long prompts, averaging around 10K tokens
Multi-turn conversations
Peak load above 1,000 requests per minute
How BlazeInfer solved it
Full-precision quality
The customer runs the BF16 version of Gemma 4 31B on BlazeInfer. Compared with quantised versions from other providers, user adoption in the app was higher.
High KV-cache reuse
Multi-turn prompts share most of their context from turn to turn. BlazeInfer's KV-cache reuse lowered the effective price, and our team worked with the app developers to raise prompt cache hit rates.
Routing and scaling
Efficient routing and scaling optimisations handled peaks above 1,000 RPM while keeping P95 end-to-end latency under 5 seconds.
Keep your code. Change the endpoint.
Add your BlazeInfer API key and update the base URL.
from openai import OpenAI
client = OpenAI(
api_key="BLAZE_API_KEY",
base_url="https://blaze.e2enetworks.net/v1"
)
response = client.chat.completions.create(
model="qwen/qwen3.8-27b",
messages=[
{"role": "user", "content": "Implement hello world in Python"}
]
)
BlazeInfer is built and operated by E2E Networks Limited, India's GPU cloud for AI and ML — the same infrastructure trusted by startups, enterprises and government workloads.
2009
Founded
NSE: E2E
India's first NSE-listed cloud provider
MeitY
Empanelled Cloud Service Provider
10,000+
Innovators building on E2E
One API. Infinite intelligence. Lower cost.
Get your API key and start calling BlazeInfer in minutes — no infrastructure to manage.