Featherless runs open-source AI models behind a single API, letting developers and enterprises call thousands of models without provisioning GPUs or managing servers. The platform supports over 30,000 open models across language, vision, and audio, enabling developers to deploy production-grade AI instantly , and the company describes itself as a neutral layer for AI, unaligned with any hyperscaler, chipmaker, or proprietary ecosystem . The target user is the team that wants to build on models it can audit and swap, rather than depend on a closed proprietary endpoint.
This launch marks the company's Series A. Featherless.ai secured $20 million in Series A funding, co-led by AMD Ventures and Airbus Ventures, with participation from BMW i Ventures, Kickstart Ventures, Panache Ventures, and Wavemaker Ventures . The capital will be used to scale global infrastructure, launch a dedicated marketplace for specialised open models, and deepen technical integration with diverse hardware architectures to continue driving down the cost of AI inference . The timing reflects a broader pull toward jurisdictional control and hardware diversity, with core infrastructure hosted in the US and EU and a global team across Canada, Europe, the US, Singapore, and Australia , plus a strategic collaboration with AMD that ensures the world's most popular open-source models run natively on AMD ROCm .
The company is co-founded by Eugene Cheah (CEO), Harrison Vanderbyl (CTO), and Wesley George (COO) , the same group behind RWKV, an open-source architecture designed to challenge the traditional dominance of transformers . That research lineage shapes the inference work here, and frames the Series A as a bet that open models, run on neutral infrastructure across multiple chip vendors, can hold their own against the small set of closed stacks currently absorbing most AI demand.
Comments (14)
Sign in to join the discussion.
Y
Yusra BouzidApr 30, 2026
The 'few countries, few companies, few chips' opener is doing heavy lifting in that tweet. Manifesto energy for a Series A is a choice and I respect it.
K
Kosta M.Apr 30, 2026
One API key for any open weights model is the kind of thing that sounds boring until you've spent a weekend wrangling vLLM containers.
B
Benoît LacroixApr 30, 2026
Distribution question: are you betting on devs swapping out OpenAI base URLs, or on a brand new audience that never had a key in the first place?
P
Praja IyerApr 30, 2026
hot take: the real moat here isn't the inference, it's whoever makes the long tail of huggingface models actually reachable without crying.
L
Linh TrầnApr 30, 2026
Curious about cold start times on the obscure stuff. Calling a model nobody else has loaded should be the real benchmark, not llama-3 latency.
S
Sasha PereiraApr 30, 2026
been telling people for months that open-weights infra is the next picks-and-shovels play. nice to be vindicated by someone else's term sheet.
K
Kwame TettehApr 30, 2026
concentration is gravity. open infrastructure is just whoever keeps paying the electric bill to push back on it.
O
Omar HaddadiApr 30, 2026
How's the docs situation? An any-model API lives or dies on whether I can find the right model card without spelunking through three tabs.
T
Toby WestergaardApr 30, 2026
naive question, if I have a finetune sitting on huggingface does it just work, or is there a 'we support these 400 models' list hiding somewhere?
M
Mira SolovyevaApr 30, 2026
The launch thread engagement curve on this is going to be interesting. Manifesto tweets either rip or get ratio'd, no in between.
F
Finn O'RourkeApr 30, 2026
Procurement hat on: SOC2? Data residency? If my legal team sees 'serverless inference for any open model' they're going to need a paper bag to breathe into.
C
Chiamaka EzeApr 30, 2026
RWKV in the bio and serverless inference as the day job is a delightful combo. Builds the engine and the fuel.
R
Ravi SubramanianApr 30, 2026
how big is the team behind this? every time I see 'any open model, one endpoint' I assume an army, then it turns out to be like nine people in a discord.
A
Anikó VargaApr 30, 2026
would love to see published throughput numbers across model sizes before I migrate anything. trust but verify and all that.
Work With Us
Book an intro call to explore our suite of services. We might have a waitlist.
Featherless AI
One API key. Instant access.
One API key. Instant access.
Tags
TLVC Rating
Community Rating
No ratings yet
to rate this launch.
Featherless runs open-source AI models behind a single API, letting developers and enterprises call thousands of models without provisioning GPUs or managing servers. The platform supports over 30,000 open models across language, vision, and audio, enabling developers to deploy production-grade AI instantly , and the company describes itself as a neutral layer for AI, unaligned with any hyperscaler, chipmaker, or proprietary ecosystem . The target user is the team that wants to build on models it can audit and swap, rather than depend on a closed proprietary endpoint. This launch marks the company's Series A. Featherless.ai secured $20 million in Series A funding, co-led by AMD Ventures and Airbus Ventures, with participation from BMW i Ventures, Kickstart Ventures, Panache Ventures, and Wavemaker Ventures . The capital will be used to scale global infrastructure, launch a dedicated marketplace for specialised open models, and deepen technical integration with diverse hardware architectures to continue driving down the cost of AI inference . The timing reflects a broader pull toward jurisdictional control and hardware diversity, with core infrastructure hosted in the US and EU and a global team across Canada, Europe, the US, Singapore, and Australia , plus a strategic collaboration with AMD that ensures the world's most popular open-source models run natively on AMD ROCm . The company is co-founded by Eugene Cheah (CEO), Harrison Vanderbyl (CTO), and Wesley George (COO) , the same group behind RWKV, an open-source architecture designed to challenge the traditional dominance of transformers . That research lineage shapes the inference work here, and frames the Series A as a bet that open models, run on neutral infrastructure across multiple chip vendors, can hold their own against the small set of closed stacks currently absorbing most AI demand.
Comments (14)
The 'few countries, few companies, few chips' opener is doing heavy lifting in that tweet. Manifesto energy for a Series A is a choice and I respect it.
One API key for any open weights model is the kind of thing that sounds boring until you've spent a weekend wrangling vLLM containers.
Distribution question: are you betting on devs swapping out OpenAI base URLs, or on a brand new audience that never had a key in the first place?
hot take: the real moat here isn't the inference, it's whoever makes the long tail of huggingface models actually reachable without crying.
Curious about cold start times on the obscure stuff. Calling a model nobody else has loaded should be the real benchmark, not llama-3 latency.
been telling people for months that open-weights infra is the next picks-and-shovels play. nice to be vindicated by someone else's term sheet.
concentration is gravity. open infrastructure is just whoever keeps paying the electric bill to push back on it.
How's the docs situation? An any-model API lives or dies on whether I can find the right model card without spelunking through three tabs.
naive question, if I have a finetune sitting on huggingface does it just work, or is there a 'we support these 400 models' list hiding somewhere?
The launch thread engagement curve on this is going to be interesting. Manifesto tweets either rip or get ratio'd, no in between.
Procurement hat on: SOC2? Data residency? If my legal team sees 'serverless inference for any open model' they're going to need a paper bag to breathe into.
RWKV in the bio and serverless inference as the day job is a delightful combo. Builds the engine and the fuel.
how big is the team behind this? every time I see 'any open model, one endpoint' I assume an army, then it turns out to be like nine people in a discord.
would love to see published throughput numbers across model sizes before I migrate anything. trust but verify and all that.