Kimi K3 Now Available via Telnyx Inference API
telnyx.comMoonshot AI released open weights for Kimi K3 today and it's live on Telnyx Inference. The architecture and Moonshot's own benchmarks are in their technical blog. This post is about running it on Telnyx.
What we are adding: K3 is now available via the Telnyx Inference API, hosted on GPUs that we own and operate.
This matters due to the size of this model. A 2.8T model needs dedicated infra to serve well. Because we own and operate the GPUs, we can contorl throughput. with no inter-provider hops and no cloud tenant, latency is minimized. We have GPUs in each of the US, EU, APAC, and MENA. Inference runs in the region you pick, with zero data retention. We do not store prompts or completions after the response returns. Because we own the infra, the per-token price reflects the cost of running the model, not the cost of renting someone else's plus their margin.
We have not benchmarked K3 ourselves yet. Moonshot's own numbers and early third-party evaluations put it at frontier level for coding and agentic work, trailing only Claude Fable 5 and GPT 5.6 Sol on aggregate. Full breakdown in their blog.
Pricing on Telnyx: $2.70/1M input tokens, $13.50/1M output tokens, $0.27/1M cached input tokens. Prompt caching enabled by default. Served via an OpenAI-compatible endpoint so you can test easily.
Model ID: [MODEL_ID] API: https://api.telnyx.com/v2/ai/chat/completions Docs: https://developers.telnyx.com/docs/inference Technical blog (Moonshot): https://www.kimi.com/blog/kimi-k3
Is this an ad?
Why is this provider specifically on the front page?
Also why is this not available over openrouter?
I don't mind it as long as the company engage in the comments. They haven't so far.
here now.
This (Telnyx) is available on TrustedRouter.com
Maybe, but how is it different than post for new Macbook?
Co-founder/CEO here. Not sure how we got to the front page, but we'll take it.
We've been trying to get on OpenRouter. Just reached out to their CEO. Would be great if people could request it as well.
We're unique in that we have B300 in USA, EU, UAE, and AUS.
Also, building out other agentic primitives like stateful actors (in beta, but will be rolled out globally in the coming weeks):
https://developers.telnyx.com/docs/edge-compute/stateful-act...
Also, if you're doing things agentically, check out AI repo:
https://github.com/team-telnyx/ai
Do you guys provide ZDR? We’ve been waiting for someone to offer inference in Sydney on decent GPUs for a while but we’ve got strict no logging or retention requirements.
Also available from Nebius, via Cortecs: https://cortecs.ai/detailedServerlessView/kimi-k3
€2.693/M input €13.464/M output Surprisingly, cache is not mentioned
Upd: Tensorix joined the fray, with the same prices, with cache at €0.673/M read
10% cheaper than official! Let the inference pricing wars begin!
I'm really curious how far and how fast prices will drop (if at all)!
In my first interaction ("hi there kimi k3!"), Kimi K3 identified twice out of three times as Claude:
> Hi there! Quick note — I'm actually Claude, made by Anthropic, not Kimi. But no worries!
https://imgur.com/a/jqpc2Jc
and
> Just a quick heads-up — I'm Claude, made by Anthropic, not Kimi! Kimi is a different AI assistant (made by Moonshot AI), so it looks like there might be a little mix-up.
https://imgur.com/a/AKxeysH
Bootleg AI
if this qualifies bootleg, point me at non-bootleg frontier ai.
Is Claude bootleg AI? https://www.reddit.com/r/OpenAI/comments/1e34tkr/why_is_clau...
That’s how a smoking gone turns into a water pistol
when open source models are banned maybe i will become a cartel kingpin smuggling open source weights into the usa. find a sufficiently shifty street corner, 'what do you need', "kimi". are you familiar with my product? pure mxfp4, $500 per TB.
(joke and walter mitty, i know the government reads my messages)
Models don't have an inherent identity. It should be obvious by now that every models trains on public AI chat session transcripts. I've seen Claude say it's Qwen.
This happens with other models too - Gemini often identifies as ChatGPT for me, confusing many a debugging attempt
It's distillations all the way down!
How many times do people need to point out that every model has this behavior until this stops being posted?
FWIW Claude sometimes identifies as DeepSeek when asked in Chinese: https://x.com/stevibe/status/2026227392076018101
I remember when Claude identified as ChatGPT a long time ago. It proves nothing else than that there is a lot of training material on the internet with Claude as the AI.
I don't understand why labs don't include correct model name in the training process. Almost nobody seems to be willing to tell their models who they actually are.
If the model works correctly, you don't need to include the model name in the training process. You just add in the system prompt "You are X, trained by Y" and the model will claim to be that. That also allows users to "white-label" the outputs sort to say.
This approach is basically how all "models know who they are" (they don't actually typically "know" that at all), it's just a system prompt instruction in the platform you use.
But I would expect they would at least use post-training to discourage identifying as the wrong company.
That only makes sense if you don't sell access to the model for 3rd party product development. One of those products is chat personas/agents. The company integrating the model will want it's own identity.
You can put anything in the prompt. And if you have bare model that you don't give any prompt from the start and want to find out which model you are talking to, you are out of luck.
I think Qwen teaches its models that they are Qwen. Most others don't bother.
Because they do not know the name of the model before they train it. There is also distillation, where multiple models will be trained from a larger one. E.G. Sonnet was promoted to Opus at one point after it surpassed expectations.
Could be on purpose to disguise as a US made model
The meme/trope of China copying everything really keeps playing into itself
Models don't know who they are. Stop asking this question to the model as a source of truth.
Why are you guys not on Openrouter? I assume you'd get way more volume that way no?
Or does openrouter have like a specific contract you have to sign with them and requirements or smth? https://openrouter.ai/moonshotai/kimi-k3#providers
Who is "you guys"? Telnyx?
Well, the OP's username is "Fiona at Telnyx"
I kinda realized my question was dumb afterwards, I think I misread the question lol thanks
This (Telnyx) is available on TrustedRouter.com
Soon, we're gonna need a router for all the different routers.
Is it also offered via a ZDR + BAA / HIPAA eligible? Would love to use it in prod
Consider publishing latency, throughput, and cost metrics under different workloads to help teams make informed decisions.
Token Type Price per 1M tokens Cached Input $0.27 Input $2.70 Output $13.50
Looks like they are first vendor to undercut in price.
I have used Telnyx for, on the opposite end of cool-ness, their fax API. Curious if their AI pricing and quality are actually competitive or if this is just a rapid pivot to try to ride the AI wave.
By 2030 or so, it will become clear that AI's impact on the economy has been no greater than the fax machine's.
It's not a rapid pivot. We've been doing AI stuff for quite some time, particularly on the voice side. Soon, you'll see more from us on the physical AI side. Our GPUs are colocated next to our mobile core, so as we roll out vision models, you'll be able to do things like agentically pilot drones with low latency.
any of these providers are HIPAA compliant?
Amazon Bedrock, but they don't have this model yet
At Telnyx, we are willing to enter into a BAA. Just hit up our sales team.
Where is it hosted ?
Atlanta, Denver, Paris, Sydney: https://developers.telnyx.com/docs/inference/models/regions
The top text says MENA as well, but the regions page doesn't mention it.
Thanks !
Very cool. What are your throughput and latency like?
That is not going to be cheap for long if K3 prices does the same as GLM5.2 prices. If nothing else open weight models are great to get providers to compeete on price.
what quantization? FP4?
The model is native mxfp4 w/ mxfp8 activations via QAT.
I have some morbid curiousity about what K3 would be like in an FP16 or FP32 model.
Jevon's paradox depends on how cheap the tokens get as the price of tokens get driven to zero as intelligence gets better and cheaper.
Well this is a very large and expensive model. Try something like Qwen-7B if you want ultra-cheap.
Larger than Mythos? How is that?
I thought Mythos was rumored to be around 10tn parameters? Kimi k3 is 2.8tn.
I think telnyx is a good product, with the only stain to its name being the supply chain attack on their python library.
But I don't feel like providing inference is a professional move, it feels like out of scope for a telephony IaaS company. Feels like a FOMO moment where a reputation of years is crashed in a couple of weekends of being drawn into a fad.
And the fact that it's a chinese model doesn't quite help? I guess it's on brand with the 'cheap' pay as you go brand telnyx might already be associated to.
But more so it reads like Telnyx is trying to 'jump' into the trend of the 'open weights' discussion to compete with closed source incumbents. But we are at the tail end of the boom, anti ai sentiment is ever growing, customers now despise AI, especially in support channels, which is presumably the hook that Telnyx would have into 'AI'(LLMs). At this stage any company or individual that tries to join into the buzzword fueled cycle will pay the full fixed cost reputational price, but only reap the leftover hay from when the sun shone.
AI(LLM) on support channels is essentially a decapitalization of a company/brand, the company has a reputation that customers value, and might be worth good money in the market, and by implementing AI (LLMs) on support, a lot of costs can be cut, while the brand loses value, not sustainable. And by Telnyx (or any B2B company)asking their clients to participate in this decapitalization move, they essentially gamble their reputation as well, albeit with better odds as shovel sellers.
Large profitable corporations with very poor customer support have existed long before AI, so why should it be different? Using AI to provide poor customer service is just an implementation detail.
Companies changing their scope or opening side products is not at all unusual. Microsoft is Windows, what's this Xbox thing? Google is search, why do they have email? Y Combinator is a startup accelerator, why'd they make their own Reddit?
I think Microsoft is still trying and failing to figure out what this X-Box thing is.
Yeah, the pypi thing sucked, and we've taken the necessary measures to prevent something like that from happening again.
We're moving past "telephony" and focusing on building more agentic primitives at the telecom edge.
This isn't a recent jump. We've been on this path for some time.
Disagree with you on the customer support side and re: decapitalization in general. People will want to talk to competent bots. We have people literally prompting our AI agents (which are capable of doing troubleshooting) in our shared Slack channels. You simply cannot beat the speed with which they can provide a quality response.
We remain committed to delivering the highest quality product at the lowest possible price point - across all of our product lines.
Telnyx was cool until they started demanding KYC. I would use them for burner phone numbers until they started saying they needed my government ID. Fuck that.
OVH same thing. Tried to buy a VPS from them some years back and they said no VPS unless I provided ID. Would not refund me. Tried to dispute but my bank just gave me a credit instead.
There were new regulations passed to combat robocalls that forced these companies to tighten up.
10DLC in North America
You can go on telegram and pay someone $10 to do the KYC for you.
hi nsa
Seems like a good riddance, telnyx is for business usecases, not for personal use.
Such use would be incompatible because it would lure in fraud, which would ruin the reputation of shared comms resources like ip blocks, phone number blocks, etc..
If you are doing real business, you are providing KYC as a daily matter in procurement, thereby protecting consumers from Sybil scum.
> thereby protecting customers from Sybil scum
No? Firstly you can't do a Sybil attack when each identity costs actual resources, and secondly you can still have lots of phone numbers, they just know who each one belongs to.
We've been fighting this battle with the FCC. They tried fining us (and then dropped it), but they continue to push the issue via NPRMs. We generally think KYC is ineffective due to ID mules. It's obviously a privacy issue as well.
That’s huge
is not compatible with
since you asserted something false and bizarre, how about telling us what is your markup?
I don't get it, what's the contradiction supposed to be?
Why are those incompatible? Pricing is 10% under moonshot, and pure infra providers also want money
They are selling it even cheaper than Moonshot AI. Why are you so sure they are lying?
https://newsletter.semianalysis.com/p/ai-value-capture-the-s...
according to Semianalysis those prices would be far above actual costs.
**reflects** the cost...
Not the **literal** cost
If they are indeed far above actual costs, then surely price discovery will be done by the overall market in due course.
They're not saying anything about Anthropic serving costs in that quote, just calculating what MTok price is for running agents. Next sentence after your quote:
> Agentic workloads have extremely high input-to-output ratios (our Claude Code usage has a ratio of about 300:1) and high cache hit rates (90%+). Because cached input tokens only cost $0.50/MTok, most of the tokens end up in the cheapest tier.
90% cache hit input blend: 0.9 * $0.5 + 0.1 * $5 = $0.95 per MTok.
300:1 input/output blend: (300 * $0.95 + $25) / 301 = $1.03 per MTok.
They don't say exact cache hit rate they calculated for ("90%+"), so close enough.
Please correct me if I'm incorrect, but it seems to me those numbers are describing a fairly different situation to this one. Anthropic serving their own model to their own users at that scale gets cache hit rates and machine utilisation that someone standing up another company's 2.8T model in four regions isn't going to get, and this thing needs 64 accelerators minimum before it will run at all, so a rack sitting mostly idle through a quiet hour costs the same as a busy one. The margin figure quoted is also just the price against the cost of producing the tokens, it doesn't have buying the hardware in it, or depreciation, or maintenance, or the money they'd have made renting those machines out instead, which with rental prices up 40% since October isn't nothing.
I largely agree things are overpriced, I just don't think that article is the right basis for saying it about this one.