Your traffic. Your deployment.
Dedicated capacity, agreed data processing and retention, and an inference stack optimised for your workload alone.
How we protect your data WORKLOAD-TUNED AI INFERENCE
Inference,
shaped by
your workload.↗
Private, fast and affordable inference.
Tuned to the traffic you actually serve.
A custom KV backend and our optimised DSpark speculative decoding turn the patterns in your workload into more tokens per GPU, and fewer dollars per token.
✓ Private. Fast. Affordable.
PC—01
SHAPED BY YOUR WORKLOAD
SHAPED BY YOUR WORKLOAD
✳
Custom KV backendTiered HBM · DRAM · NVMe
↗ DRAG TO EXPLORE
STANDARD SERVING
Output tokens / s / GPU760×1.0KV hit 22%τ 1.0TTFT 1.9 s
CREATIVE WRITING · Qwen3.8-Flash-Next · 4× H100×0.0more output tokens per GPUQwen3.8-Flash-Next · 4× H100
−0%cost per million output tokensSame GPUs · same latency target
0.0tokens accepted per verificationDSpark, tuned to the workload
0%of prompt tokens served from cacheTiered HBM · DRAM · NVMe
01 / THE AVERAGE WORKLOAD DOESN’T EXIST
Built for everyone.
Tuned for no one.
Public APIs and stock serving stacks are tuned for the average request. But your prompts repeat, your contexts overlap and your outputs follow habits. We tune the whole inference path for your traffic, and measure the difference on your own workload.
Tokens arrive in blocks.
DSpark drafts several tokens at once and the target model verifies them in a single pass. A KV backend that remembers your context cuts the wait for the first token.
Inside the technologyThroughput is the price.
When every GPU produces more tokens per second, every token costs less. Same hardware, same latency target, a smaller bill.
Your token economics02 / DSPARK, OPTIMISED IN-HOUSE
Predict what your
workload says next.
Standard decoding produces one token per forward pass. DSpark drafts a block and the target model verifies it at once. Ours is optimised for your workload, so more of every draft is accepted, with no change to the output.
BASELINE
Standard decoding
- FORWARD PASSES
- 0
- TOKENS / PASS
- —
- vs STANDARD
- …
PUBLISHED DRAFTER
Stock DSpark
- FORWARD PASSES
- 0
- TOKENS / PASS
- —
- vs STANDARD
- …
WORKLOAD-TUNED
Our DSpark
- FORWARD PASSES
- 0
- TOKENS / PASS
- —
- vs STANDARD
- …
Standard decoding: about 1 tokens per forward pass. Stock DSpark: about 2.4 tokens per forward pass. Our DSpark: about 6.2 tokens per forward pass.
Token from the target model Drafted token, verified and accepted Draft awaiting verification
SPEED THAT HOLDS UNDER LOADMost speculative decoders fade as traffic grows.
Most speculative decoders fade as traffic grows.
Ours keeps its lead.
At high concurrency, GPUs are already busy and extra drafts cost more than they save. We schedule the verify budget against your concurrency and latency target, so the speed-up survives production traffic.
0.0×faster than standard decoding at 32 concurrent requests
How the scheduling works | Concurrent requests | Our DSpark | DFlash | EAGLE-3 |
|---|---|---|---|
| 1 | 5.9 | 5.1 | 1.6 |
| 4 | 5.6 | 4.5 | — |
| 8 | 5.3 | 4.5 | — |
| 16 | 4.9 | 3.9 | — |
| 32 | 4.3 | 2.8 | 1 |
03 / A KV BACKEND THAT REMEMBERS
Never compute the same
context twice.
Agents, assistants and document pipelines resend the same prompts, tools and histories every turn. Our KV backend keeps that context across GPU memory, host memory and local NVMe, including for hybrid-attention models where stock caches miss.
ONE SESSION · AGENT CONVERSATION, 8 TURNS
Recompute the whole context Reuse cached KV, prefill only new tokens
6K-token system and tool prefix · about 5K tokens added per turn · 12K tokens/s prefill · 400K tokens/s KV load from the cache tiers.
UNDER REAL AGENTIC LOAD
0× faster to the first token.
Average time to first token across 32 concurrent coding-agent sessions with 100K-token contexts.
04 / MEASURED ON YOUR WORKLOAD
Every workload has a shape.
Tuning to it changes everything.
The same models on the same GPUs, at the same latency target. Only the serving stack changes.
WHERE STOCK METHODS STRUGGLE MOSTThe harder the workload,
The harder the workload,
the bigger the gain.
Creative writing defeats generic speculative decoding and prefix caching: every reply is new prose, and readers pause long enough between turns for ordinary caches to forget the story. It is also where optimising for the workload pays off most.
DSPARK · TOKENS ACCEPTED PER VERIFICATION Stock DSpark Ours
Creative writingBIGGEST GAIN
2.46.2
×2.6Support chat
3.75.9
×1.6Knowledge assistant
3.95.6
×1.4Coding agent
5.58.1
×1.5Document extraction
6.210.2
×1.6KV BACKEND · PROMPT TOKENS FROM CACHE Standard prefix cache Ours
Creative writingBIGGEST GAIN
22%91%
×4.1Support chat
61%93%
×1.5Knowledge assistant
35%78%
×2.2Coding agent
64%96%
×1.5Document extraction
48%89%
×1.9WORKLOAD PROFILE
Creative writing
Long stories and character dialogue at high temperature. Every reply is new prose, and readers pause long enough between turns for ordinary caches to forget the story.
- MODEL
- Qwen3.8-Flash-Next
- HARDWARE
- 4× H100
- LATENCY TARGET
- ≥ 40 tok/s per user
×0.0output throughput per GPU
STANDARD
760
TUNED
tokens / s / GPU0
STANDARD
$1.46
TUNED
USD · at $3.99/GPU-hour$0.00
STANDARD
1.90
TUNED
seconds0.00
STANDARD
22
TUNED
% of prompt tokens0
05 / BETTER ECONOMICS
Less spend.
More tokens.
See what higher throughput does to your cost per token and to the number of GPUs you need.
THROUGHPUT IS THE PRICEMore tokens per GPU.
More tokens per GPU.
Fewer dollars per token.
Cost per token is GPU time divided by the tokens it produces. Raise throughput on your workload and every token gets cheaper, on the same hardware.
GPUS NEEDED14 → 3
COST PER 1M OUTPUT TOKENSUSD
Standard serving$0
Tuned stack$0
Saved every year$0
Throughput uplift ×6.8 measured on a creative writing workload. Standard serving: SGLang without speculative decoding, 8K in / 1K out, 64 concurrent requests (SemiAnalysis InferenceX). GPU counts assume 60% average utilisation. List prices as published on 24 September 2026. Public APIs for open-weight models can list lower prices; they are shared services, not tuned to your workload.
06 / YOUR APPLICATION. OUR OPERATIONS.
Private by default.
Tuned for you alone.
Connect through an API. Perchy runs the tuned stack on capacity reserved for your workload, and keeps server administration.
MANAGED BY PERCHY. CONNECTED BY API.
An API for your applications
Send authenticated requests and receive model outputs. Perchy runs the models and the tuned inference stack, including server administration and maintenance.
- Perchy-operated infrastructure
- Authenticated API access
- Agreed data processing and retention
For teams building AI features through a managed API.
Discuss this service YOUR APPLICATION → PERCHY API
⌘Your applicationPRODUCT / WORKFLOW / API
Authenticated API requests
PERCHY-OPERATED INFRASTRUCTURE
Perchy gateway
WORKER 01
WORKER 02
WORKER 03
MANAGED MODEL INFERENCE
Perchy administers the servers · customers connect through the API
A CLEAR SERVICE BOUNDARY
Inference, through an API.
Perchy provides managed AI inference through authenticated APIs. Customers do not receive SSH, root, operating-system or bare-metal access. The service does not provide model-training or fine-tuning functionality. Server administration remains with Perchy.
07 / ONE STACK. ANY MODEL.
Any model.
Accelerated for your workload.
01 / LANGUAGE
HYBRID ATTENTION
Qwen3.8-Flash-Next
Recurrent state meets sparse attention. A different way to carry context through an intelligent workflow.
Customer conversations · Agentic workflows · Knowledge in context
Inside the architecture 02 / LANGUAGE
COMPRESSED SPARSE ATTENTION
DeepSeek-V4.1-Flash
A causal encoder–decoder, compressed sparse attention and conditional memory. Explore the architecture behind the next token.
Long-context reasoning · Coding and agents · Multimodal knowledge
Inside the architecture 03 / LANGUAGE
SPARSE + LINEAR ATTENTION
GLM-5.3-Flash
Kimi Delta Attention and KPool-DSA share the work: recurrent state for continuity, selective attention for recall.
Engineering workflows · Long-context recall · Visual understanding
Inside the architecture 04 / IMAGE
SINGLE-STREAM FLOW MODEL
Ideogram 4.0
Language and image tokens meet inside a single-stream diffusion transformer, turning latent noise into deliberate composition.
Product imagery · Typography-led creative · Creative iteration
Inside the architecture 05 / VIDEO
JOINT AUDIO–VIDEO TRANSFORMER
MiniMax H3
A joint audio–video latent sequence. One single-stream transformer. A generation path designed to keep sound and motion together.
Product stories · Creative exploration · Content pipelines
Inside the architecture Model examples illustrate managed inference options. Availability, licences and performance are confirmed in your proposal. Model names belong to their respective owners; listing does not imply endorsement or partnership.
One benchmark.
A whole new cost curve.