Qwen-Image
- Zero Data Retention
- No Rate Limits
- No Contracts
Blazing Fast Agentic Inference
— One Endpoint
Run workloads on SOTA open-weight models, custom fine-tunes, with batch jobs — autoscaled, observable, failover resilient, hosted in India.
- LIVE
80
ms median
9,864
tok/sec
99.99
- [ PLATFORM ]
Performance, Control, Compliance.
Built for scale.
Built for teams to scale production-grade inference — within the SLA boundary.
01
.
- Elastic Inference
- LIVE
Serverless & autoscale on demand. Failover built in. Zero cold-start concerns.
Concurrent requests · last 60s 39,988
-60s -45s -30s -15s now
Cold start
— ms
Autoscale
0 ➔ ∞
Failover
Multi-AZ
02
.
- Developer First
Scale to billions of tokens in hours.
No rate limits. No quotas. Ship the moment your workload spikes.
03
.
- Zero Data Retention
Prompts and outputs never persist.
No logs. No storage. Nothing leaves memory after the response.
stdout ➔
79920 d3e28 5f36a
/dev/null04
.
- Full-Stack Observability
Latency, throughput, cost, and failure rates — visible at every layer.
From the gateway to the GPU, every hop is instrumented and queryable.
Latency
Throughput
Cost
Failures
GPU
- [ Every modality ]
SOTA models for text, image, and video
One API, every modality. Reason, generate, transcribe, and edit
— across the best open-weight models in each category, all
.running on the same elastic infra.
- streaming · 312 tok/s
Namaste — serverless inference for the agentic
era
era
- Batched
- streaming
- confidential
- Deploy Llama
- Qwen
- Sarvam
- Run in Mumbai or
@registry-org
128K context
- live
- Large Language
Reason, write, call tools.
Frontier open-source LLMs, tuned for lowest TTFT and highest throughput. Streaming, structured outputs, and native tool-calling — out of the box.
Kimi K2.6
GLM 5.1
DeepSeek V4 PRO
Flux · 1024×1024 · 2.1s
- Image Generation
Generate, edit, upscale.
Flux, Qwen, and Stable Diffusion 3 — all running on dedicated image pods. Superfast generation with zero storage.
Flux Klein
- Large Language
Reason, write, call tools.
Frontier open-source LLMs, tuned for lowest TTFT and highest throughput. Streaming, structured outputs, and native tool-calling — out of the box.
Kimi K2.6
GLM 5.1
DeepSeek V4 PRO
- [ MODELS ]
Every open - weight model. One endpoint.
Hot-swap between Llama, Qwen, DeepSeek, Mistral, Gemma, and Sarvam. Bring your own checkpoint or deploy directly from Hugging Face with a single command.
Deepseek V4
deepseek-ai/DeepSeek-V4-Pro 40 tok/s
Kimi 2.6
moonshotai/Kimi-2.6 50 tok/s
GLM 5.1
THUDM/GLM-5.1 70 tok/s
Minimax 2.7
minimaxai/Minimax-2.7 120 tok/s
DeepSeek V4 Flash
deepseek-ai/DeepSeek-V4-Flash 200 tok/s
Qwen 3.5 35B
Qwen/Qwen3.5-35B-Instruct 150 tok/s
Deepseek V4
deepseek-ai/DeepSeek-V4-Pro THROUGHPUT
40 tok/s
CONTEXT
128K
INPUT / 1M TOK
₹60
OUTPUT / 1M TOK
₹180
> Reason through a 4-step trade settlement reconciliation.
← Step 1 · match TRD-2451 → bank ref BR-9821 (amount Δ ₹0). Step 2 · flag broker fee mismatch on TRD-2452 …
OpenAI-compatible
streaming
JSON mode
FP8 quant
tool-calling
Kimi 2.6
moonshotai/Kimi-2.6 THROUGHPUT
50 tok/s
CONTEXT
200K
INPUT / 1M TOK
₹50
OUTPUT / 1M TOK
₹150
> Summarise this 80-page RFP into three bullet points.
- Vendor must support DPDP-aligned residency
- 99.95% SLA with 30-min credit clause
- Decision by Q3
OpenAI-compatible
streaming
JSON mode
BF16 quant
tool-calling
vision
GLM 5.1
THUDM/GLM-5.1 THROUGHPUT
70 tok/s
CONTEXT
128K
INPUT / 1M TOK
₹35
OUTPUT / 1M TOK
₹105
> Generate a JSON schema for a kirana invoice.
← { "type":"object", "properties":{ "items":[…], "gst":{ "type":"number" } } }
OpenAI-compatible
streaming
JSON mode
FP8 quant
tool-calling
Minimax 2.7
minimaxai/Minimax-2.7 THROUGHPUT
120 tok/s
CONTEXT
1M
INPUT / 1M TOK
₹28
OUTPUT / 1M TOK
₹84
> Translate this 30-min meeting transcript and tag action items.
- Owner: Priya · Action: ship payments hot-fix by Fri
- Owner: Rahul · Action: align with legal on DPDP scope
OpenAI-compatible
streaming
JSON mode
FP8 quant
tool-calling
vision
Deepseek V4 Flash
deepseek-ai/DeepSeek-V-Flash THROUGHPUT
200 tok/s
CONTEXT
64K
INPUT / 1M TOK
₹12
OUTPUT / 1M TOK
₹36
> Write a Python function to dedupe rows by fuzzy match.
←
def dedupe_fuzzy(rows, key, threshold=0.92):
# rapidfuzz.process.extract → cluster → keep canonical
OpenAI-compatible
streaming
JSON mode
INT8 quant
tool-calling
Qwen 3.5 35B
Qwen/Qwen3.5-35B-Instruct THROUGHPUT
150 tok/s
CONTEXT
128K
INPUT / 1M TOK
₹18
OUTPUT / 1M TOK
₹54
> Classify this UPI complaint by type and urgency.
←
category: refund_pending · severity: medium · sla: 24h
OpenAI-compatible
streaming
JSON mode
FP8 quant
tool-calling
vision
** tok/sec on shared endpoints is subject to differ based on real-time traffic. Opt for dedicated endpoints for guaranteed performance.
- [ Observability ]
Every request, instrumented.
Every layer, visible.
samaira.ai/observability
- live · last 5m
Requests / sec
30,418
▲4.2% vs 5m
Tokens / sec
1,415,273
▲6.1%
Error rate
0.05%
— stable
Cost / 1M tok
$0.66
▼1.8%
Request Latency
- p50
- p90
- p99
window: 5m · 1s buckets
Success rate
99.93%
Failover events
417
Retries
1703
GPU utilization
78%
- [ The Infinite Architecture ]
Supercharge your AI agents with compliance and infinite scale.
Frontier inference, inside the boundary, pay in INR.
- India Compliant · DPDP aligned.
- EU Compliant · GDPR-ready.
- Zero Logs Policy
- [ On-Prem ]
Run the Samaira Stack On Prem
End-to-end GPU orchestration, inside your infrastructure.
End-to-End GPU Orchestration
Full-stack GPU cluster management — provisioning, scheduling, and scaling on your own hardware.
Agentic Tuner
AI-driven auto-tuner that maximizes GPU utilization and inference performance for your workload mix.
Agentic Sandbox
Secure execution environment for multi-step agent workflows and tool-use chains on private infra.
TEE Support & Observability
Hardware-level trust with Trusted Execution Environments plus full-stack observability built in.
- [ Roadmap ]
What's coming next.
Coming Soon
TEE Support
Confidential compute for workload isolation and hardware-level trust. Encryption in use, attestation by default.
- Q3 · Private alpha
Coming Soon
Dedicated Endpoints
Reserved capacity, custom scaling policies, and endpoint-level monitoring for predictable production workloads.
- Q2 · Closed beta
Coming Soon
Agentic Sandbox
Secure, sandboxed execution environment for agent workflows, tool use, and multi-step reasoning chains.
- Q4 · Research preview
Enterprise AI inference,
built for India.
Secure, fast, and fully visible. Talk to us about bringing your
inference workloads inside the boundary.
$ curl https://inference.samaira.ai/openai/v1/chat/completions \
-H "Authorization: Bearer $SAMAIRA_KEY" \
-H "Content-Type: application/json" \
-d '{"model":"MiniMaxAI/MiniMax-M2.7","messages":[{"role":"user","content":"Hello, India."}],"stream":false}'
{ "id": "chatcmpl-RGEzCmIB...", "object": "chat.completion", "choices": [{ "message": { "role": "assistant", "content": "Namaste! How can I help you today?" } }], "usage": { "prompt_tokens": 44, "completion_tokens": 12, "total_tokens": 56 } }