Parasail

Inference that scales with you.

[Parasail @parasail_io]

We're #1 on Kimi K3 (max) in @ArtificialAnlys's provider benchmark.

186 output tokens/sec and 12.75s to first answer token, across ten providers.

At $2.31 per 1M blended, matching the lowest price in the field. Fastest of the ten, priced with the cheapest.

Live now with the full 1.05M context.

https://t.co/0NCS0Eem5t

[Parasail @parasail_io]

Rate limits and regional outages are routing problems.

A production LLM gateway needs to direct traffic to healthy capacity and reroute 429s and 5xxs.

Our guide covers the architecture:
https://t.co/vMlqX1P7ZQ

[Parasail @parasail_io]

Production inference traffic rarely holds steady.

Dedicated capacity protects latency and reliability, but GPUs sized for peak demand can sit idle for hours.

Parasail actively right-sizes deployments as traffic changes, so customers keep the performance they need without paying for idle capacity.

Read more: https://t.co/UhMyXvlsoe

[Parasail @parasail_io]

Parasail is sponsoring Ray Summit 2026 in San Francisco, Aug. 24–26.

We’ll be there for the first vLLM Conference and technical sessions on production AI systems.

Working through latency, capacity, or cost for OSS inference? Come find us.

https://t.co/fihMCxbjX1

[Parasail @parasail_io]

To hit a sub-600ms p99 target on one LLM workload, we used two high-end GPUs.

After tuning for its request patterns, a single GPU delivered nearly the same latency and now meets the target in production.

How we tested the full request path:
https://t.co/3MFUvkeFnM

[Kai Huang @KaiHuang]

@parasail_io just topped the performance charts for Kimi K3! Incredibly proud to be part of the team here.

[Sam @Sam_Awrabi]

Congrats @parasail_io : fastest and lowest latency infrence endpoints in market for Kimi K3

Three new funds were allowed in the Series A @Banyan_Ventures Fund I was one of them along @KindredVentures, Touring Captial

Banyan lead the latest round congrats Parasail Team!

[Parasail @parasail_io]

Multi-region LLM deployments are a routing and failover problem before they’re a model problem.

A gateway centralizes retries, 429 handling, regional routing, and spend tracking. Extra regions don’t automatically add capacity and quota scope still matters.

Read the guide: https://t.co/4CIDHYVEhf

[Parasail @parasail_io]

Congratulations to @trygravityai on the Series A.

Gravity runs a real-time ad platform for AI applications, and every impression rides on high-performing inference.

We’re proud to be their inference partner through explosive growth from Seed through Series A. Cheers to your continued success 🥂!

[Gravity @trygravityai]

Over the past year of building Gravity, we’ve had consumer AI applications grow revenue by up to 14,722%, helped teams go from idea to raising a Series A, and most importantly, we’ve given 14.3m people free access to AI products. All by creating an Ad Network for brands to place advertisements inside of AI conversations.

We’ve raised $38m, with @lightspeedvp and Committed Capital co-leading the Series A and @caffeinatedcap leading our Seed round with participation from @BasisSet, @haystackvc, Logos Fund, @grant__gordon and more.

We’re also fortunate for the support of our incredible angels, including @parkerharris, @alexatallah, @scottbelsky, @rauchg, @ianwcrosby, Nick Marwell and @_sholtodouglas, @jgsilverman, @Zep_Haus, and many others.

Our mission is simple: Make AI free by embedding helpful and relevant advertisements into AI products so that everyone has access to frontier technology.

[Parasail @parasail_io]

Kimi K3 has been a game changer for open source models. Just a heads up though!

Many clients and SDKs are stripping reasoning tokens on /v1/chat/completions. This is roughly the correct behavior for e.g. Deepseek R1, but *it is wrong for kimi-k3*.

Kimi-K3 wants you to pass all the CoT tokens for previous messages.

If you don't do this (and, again, many clients and SDKs do this transparently without your knowledge), kimi will many-shot in-context learn not to reason regardless of your requested reasoning_effort!

So if you see kimi-k3 omitting the reasoning after a few messages, *your client is very likely broken* and Kimi will have Haiku or GPT 4.5 level intelligence.

[Parasail @parasail_io]

We're excited to partner with the @dMatrix_AI team and @nvidia on the first public announcement of a heterogenous compute solution targeted specifically for fast token generation.

As @sidsheth said “Heterogeneous compute isn't a theory anymore. It's here.”

[theCUBE @theCUBE]

#RAISESummit: Join @furrier with @dMatrixAI's Sudeep Bhoja & @sidsheth & explore their new #inference solution for fast #tokengeneration in collab with @nvidia & @parasail_io.

💡 Get more insights!
https://t.co/zDJrIjJI8G

@solidigm @NYSEWired #dMatrixCorsair #CEOInsights

[Parasail @parasail_io]

https://t.co/dJXHHiEiZE

[Parasail @parasail_io]

Kimi K3 is live on Parasail.

It is the largest open-weight model in the world (2.8T parameters), and is the most capable open model we've worked with so far.

We're excited to bring fast, compliant, and reliable production-grade access to Kimi K3 to more users with no rate limits at scale. Link to get started in thread.

[Parasail @parasail_io]

https://t.co/9rnBd1V8qH

[Parasail @parasail_io]

From talking to hundreds of AI-native startups, one of the biggest challenges we hear is scaling inference. From rate limits, to reliability issues, to runaway costs, getting inference right can be the difference between life and death.

So we're hosting a small, invite-only founders dinner Saturday, August 1st at the Agentic AI Summit in Berkeley to compare notes. Come join us for good food and an opportunity to connect and learn from peers.

Link in thread.

[Parasail @parasail_io]

We're excited to join @youdotcom @Replit @awscloud @llama_index and others for a hackathon next Friday, focused on building production-ready AI agents at the AWS builder loft. Come join us!

[You.com @youdotcom]

In 10 days, we're hosting a hackathon full of developers looking to build agents that reason over live web data, not stale training data.

We're spending a full day at AWS Builder Loft SF with @AWScloud, @CrewAIinc, @Agnoagi, @llama_index, @replit + more. (Plus $1K+ prizes!) Stay tuned to see what they ship. 🤖

[@dMatrix @dMatrix_AI]

We're excited to announce that @parasail_io is deploying d-Matrix Corsair™ inference accelerators alongside @nvidia Hopper and Blackwell GPUs to power heterogeneous AI inference.

By pairing NVIDIA GPUs for compute-intensive prefill with Corsair for latency-sensitive decode, Parasail can deliver up to 10x faster interactive token generation with improved cost-performance and energy efficiency.
The future of AI inference is heterogeneous.

Read more: https://t.co/rXe64Snyio

#AI #Inference #LLM #GenAI #NVIDIA #AIInfrastructure #dMatrix @Parasail_io

[Parasail @parasail_io]

https://t.co/yP8nuMifDz

[Parasail @parasail_io]

We’re excited to partner with @dMatrix_AI and @nvidia to deliver up to 10x faster interactive token generation.

“Parasail is proving what we’ve believed from the start: the future of inference is heterogeneous,” said @sidsheth, founder and CEO of d-Matrix. By deploying Corsair alongside their NVIDIA fleet, Parasail is embracing a heterogeneous inference approach designed to optimize each phase of the workload across multiple accelerators.”

D-Matrix’s incredibly fast SRAM-based inference and NVIDIA’s willingness to collaborate and open up their system makes this all possible. Link in thread.

[Parasail @parasail_io]

https://t.co/XG5Oszu3Tg

[Parasail @parasail_io]

Last month's recall of Fable was a good reminder of the supply-chain risks in adopting closed models. If a government directive or a vendor decision can pull the model your product runs on, your inference architecture has a failure point you don't control.

In his latest post, Gabriel Perácio breaks down how to build defensible AI inference infrastructure (link in thread).