Twitxy Saved

FuriosaAI

Verified account This account is verified. Learn more

@FuriosaAI

Our mission is to make AI computing sustainable, enabling everyone on Earth access to powerful AI.
furiosa.ai Joined October 2023
12 Following 1K Followers 114 Posts

Latest posts 5

Pinned post

FuriosaAI Verified account @FuriosaAI
Today, FuriosaAI is announcing our $620 million Series D funding round, with participation by the National Growth Fund of Korea alongside global infrastructure investors Infinitum Capital, Jericho Capital Asset Management L.P., QUDA Capital Partners, Cisco Investments, NeoFort Holding, and Worldline Capital, as well as continued backing from NAVER Ventures, DSC Investment, and KDB Capital. Korea Development Bank, Nextrans, Kakao Investment, and many of our early investors also participated.

This capital directly fuels manufacturing scaling to meet surging enterprise demand for RNGD. Additionally, we are accelerating the development of our third-generation accelerator for hyperscale data centers, co-developed with Broadcom.

From day one, our conviction has been that true compute efficiency requires silicon purpose-built for the workload. With 20,000 RNGD accelerators manufactured this year, doubling in 2027, and enterprise deployments scaling globally, we are focused on delivering real performance and sustainable economics to data centers worldwide.

Read the full announcement: furiosa.ai/blog/furiosaai…
0:15
FuriosaAI Verified account @FuriosaAI
Furiosa SDK 2026.4 is here.

Our latest software release for Furiosa RNGD, our purpose-built AI inference processor, brings improvements across performance, caching, multimodal serving, and model support.

Highlights include:

→ Higher token throughput: Profile-guided decode batching is now enabled by default, delivering a median 22.7% increase in ShareGPT output-token throughput in GPT-OSS-120B benchmarks across 20 matched configurations

→ Hierarchical KV caching (experimental): KV cache can now extend to host memory or be shared across machines, enabling prefix reuse in more deployment scenarios

→ Optimized multimodal inference: Improved vision encoding, image preparation, and embedding reuse bring vision-language workloads onto the standard serving scheduler and apply to all currently supported vision-language models, including Gemma 4, Qwen3-VL, and EXAONE 4.5

→ Broader model support: Expanded coverage across generation, vision-language, embedding, and reranking models

→ Simplified model deployment: Furiosa Executable Bundles (FXB) are now the standard format for compiled kernels across supported models

As AI inference scales, efficiently serving diverse workloads while making better use of available compute and memory continues to be essential to production deployment. With Furiosa RNGD in mass production, we’re continuing to advance the software and systems capabilities that help enterprises run state-of-the-art AI workloads more efficiently on existing air-cooled infrastructure.

⚡ Discover what’s new in our SDK 2026.4 blog: furiosa.ai/blog/furiosa-s…
⚙️ Explore the full release notes and upgrade guide:
developer.furiosa.ai/v2026.4.0/en/w…
1 replies 0 reposts 7 likes 1.8K views View on X
FuriosaAI Verified account @FuriosaAI
Today at #REALSummit2026 in Seoul, FuriosaAI and @samsungsds took the stage together to talk about what it takes to build AI infrastructure for token generation at scale and how our teams are working together to bring NPU-based inference to Samsung Cloud Platform (SCP).

Since we first began testing RNGD with Samsung SDS, our teams have worked closely across the stack, from model optimization and FuriosaLLM to device management, cloud-native integration, and the infrastructure needed to run NPU workloads on SCP.

Korea’s first NPUaaS, launched with RNGD on SCP in July, was an important first step. The service is already in commercial use, with enterprises using it to run AI inference more efficiently on NPU infrastructure. Our teams are continuing to expand model and service support, including AIOps, while evolving the hardware and software infrastructure needed to scale NPU-based inference.

The future is about going further.

We shared a preview of what’s next: a rack-scale system designed to connect 64 RNGDs, alongside continued development across the software stack to support larger and more diverse AI workloads.

Thank you to everyone who stopped by the SCP main booth, where we showcased RNGD and furiosa-llm running AI inference on SCP, including an NXT RNGD Server equipped with eight RNGDs, and to everyone who joined the sessions today. We’re excited to keep building with Samsung SDS and look forward to what’s next.
0 replies 3 reposts 9 likes 1.4K views View on X
FuriosaAI Verified account @FuriosaAI
We’re honored to see FuriosaAI recognized in the South Korean government's official briefing following President Lee Jae-Myung’s AI summit in San Francisco with leaders from @AnthropicAI, @OpenAI, @Broadcom, and @nvidia.

During his meeting with President Lee, Broadcom CEO Hock Tan highlighted Broadcom's ongoing collaboration with Furiosa on next-generation AI semiconductor development.

Collaboration is critical for South Korea and the global tech ecosystem as we build the next chapter of AI. Furiosa is proud to help strengthen Korea’s position in this vital area, and we look forward to continuing our work with leading global partners to advance the future of AI acceleration.

Read the official government briefing (Korean): korea.kr/news/policyNew…
FuriosaAI Verified account @FuriosaAI
Today we're proud to announce Furiosa SDK 2026.3, a major milestone that represents months of engineering innovation and lays the foundation for the next generation of AI model support 🚀

Behind the scenes, our team has been focused on a challenge that's becoming increasingly important: how do you keep up as new AI model architectures emerge every few months?

This release isn't just about supporting another model. It's about creating a more scalable way to support future models.

At the core of SDK 2026.3 is TCL (Tensor Contraction Language), a new kernel framework that shifts model enablement from months of architecture-specific kernel engineering to a modular, reusable foundation. The result: faster bring-up of new models, broader architecture coverage, and a dramatically more scalable path for supporting the rapidly evolving open-model ecosystem on RNGD.

SDK 2026.3 introduces:

⚡ First-class multimodal serving with Qwen3-VL

⚡ Support for major MoE families, including gpt-oss, Solar-Open, Qwen3 MoE, and K-EXAONE

⚡ FXB (Furiosa Executable Bundle), a portable compiled artifact format for zero-recompilation deployment

⚡An opt-in overlap scheduler that keeps the NPU better utilized by overlapping host scheduling with device execution

⚡ Scoring-based Data Parallel routing for improved cache locality and load balancing

For developers, this means faster access to emerging model architectures.

For enterprises, it means AI infrastructure that can adapt more quickly as the ecosystem evolves, without sacrificing performance or operational simplicity.

—
📑 Read the blog for a deeper dive: furiosa.ai/blog/furiosa-s…

📋 Explore the full release notes: developer.furiosa.ai/v2026.3.0/en/w…