lee retweeted
For our after-hours booking flow, v4 Turbo is finally fast enough that it doesn’t feel like you’re talking to a recording.
Having one consistent voice throughout the whole call is exactly what we’ve been looking for.
I think this could make a noticeable difference in completion once we roll it out.
Introducing Eleven v4 and Eleven v4 Turbo, our fastest and most emotive voice models yet.
Ranked #1 by Artificial Analysis.
16
110
119
11,379
lee retweeted
We’re moving our IVR over to Eleven v4 Turbo, and the response time already makes a huge difference.
It doesn’t have that obvious “this is a recording” feel anymore, even on the first menu.
I’m expecting callers to stick around and engage more once it’s live.
Introducing Eleven v4 and Eleven v4 Turbo, our fastest and most emotive voice models yet.
Ranked #1 by Artificial Analysis.
14
123
127
10,585
Our sales dialer runs on @elevenlabs, and honestly, the latency on v4 Turbo is the biggest deal for us.
Reps can keep the same energy throughout the call, and people can actually respond in real time without those awkward pauses.
I think this is going to make a real difference to our numbers.
Introducing Eleven v4 and Eleven v4 Turbo, our fastest and most emotive voice models yet.
Ranked #1 by Artificial Analysis.
15
91
99
10,603
lee retweeted
We’ve been stuck dealing with latency in enterprise voice for the past year.
Eleven v4 Turbo finally feels like we can have both speed and expressiveness without compromising on either.
We’re moving quickly to get this in front of customers.
Introducing Eleven v4 and Eleven v4 Turbo, our fastest and most emotive voice models yet.
Ranked #1 by Artificial Analysis.
15
113
124
17,959
There are two stories in this launch: the model that shipped, and the system used to build it. I think the second one deserves more attention.
Introducing Naive-N0.5-Flash: Building Frontier AI with AI
🔹 309B MoE, 15.5B active: top-tier in coding, leading in AI R&D.
🔹 Native 1M context, no full-attention layers (hybrid SWA + DSA).
🔹 Inference runtime built by AI: up to 2,000 tok/s in Ultrafast mode.
Weights are open today under MIT license.
🔗 Tech blog: naive.ai/en/research/
🤗 Hugging Face: huggingface.co/NaiveAI/Naive…
💻 GitHub: github.com/NaiveAI-Labs/Naiv…
🌐 naive.ai/en/
14
109
123
10,372
The benchmark numbers are nice, but “AI helped build the AI” is the line that stuck with me.
Introducing Naive-N0.5-Flash: Building Frontier AI with AI
🔹 309B MoE, 15.5B active: top-tier in coding, leading in AI R&D.
🔹 Native 1M context, no full-attention layers (hybrid SWA + DSA).
🔹 Inference runtime built by AI: up to 2,000 tok/s in Ultrafast mode.
Weights are open today under MIT license.
🔗 Tech blog: naive.ai/en/research/
🤗 Hugging Face: huggingface.co/NaiveAI/Naive…
💻 GitHub: github.com/NaiveAI-Labs/Naiv…
🌐 naive.ai/en/
13
113
194
10,215
lee retweeted
We used to optimize the model-building process with better tools. Now we're starting to use models to optimize the model-building process itself.
Introducing Naive-N0.5-Flash: Building Frontier AI with AI
🔹 309B MoE, 15.5B active: top-tier in coding, leading in AI R&D.
🔹 Native 1M context, no full-attention layers (hybrid SWA + DSA).
🔹 Inference runtime built by AI: up to 2,000 tok/s in Ultrafast mode.
Weights are open today under MIT license.
🔗 Tech blog: naive.ai/en/research/
🤗 Hugging Face: huggingface.co/NaiveAI/Naive…
💻 GitHub: github.com/NaiveAI-Labs/Naiv…
🌐 naive.ai/en/
14
111
160
9,966
lee retweeted
New lab, first model, open weights, and an AI-assisted research loop. The model is only the first data point.
Introducing Naive-N0.5-Flash: Building Frontier AI with AI
🔹 309B MoE, 15.5B active: top-tier in coding, leading in AI R&D.
🔹 Native 1M context, no full-attention layers (hybrid SWA + DSA).
🔹 Inference runtime built by AI: up to 2,000 tok/s in Ultrafast mode.
Weights are open today under MIT license.
🔗 Tech blog: naive.ai/en/research/
🤗 Hugging Face: huggingface.co/NaiveAI/Naive…
💻 GitHub: github.com/NaiveAI-Labs/Naiv…
🌐 naive.ai/en/
14
109
119
10,522
lee retweeted
“AI for AI” sounds abstract until it starts designing, running, and analyzing the experiments behind the model itself
Introducing Naive-N0.5-Flash: Building Frontier AI with AI
🔹 309B MoE, 15.5B active: top-tier in coding, leading in AI R&D.
🔹 Native 1M context, no full-attention layers (hybrid SWA + DSA).
🔹 Inference runtime built by AI: up to 2,000 tok/s in Ultrafast mode.
Weights are open today under MIT license.
🔗 Tech blog: naive.ai/en/research/
🤗 Hugging Face: huggingface.co/NaiveAI/Naive…
💻 GitHub: github.com/NaiveAI-Labs/Naiv…
🌐 naive.ai/en/
14
109
120
10,437
lee retweeted
The model is one thing. Having AI help build the runtime it runs on is a much more interesting experiment.
Introducing Naive-N0.5-Flash: Building Frontier AI with AI
🔹 309B MoE, 15.5B active: top-tier in coding, leading in AI R&D.
🔹 Native 1M context, no full-attention layers (hybrid SWA + DSA).
🔹 Inference runtime built by AI: up to 2,000 tok/s in Ultrafast mode.
Weights are open today under MIT license.
🔗 Tech blog: naive.ai/en/research/
🤗 Hugging Face: huggingface.co/NaiveAI/Naive…
💻 GitHub: github.com/NaiveAI-Labs/Naiv…
🌐 naive.ai/en/
14
110
124
8,668
lee retweeted
The interesting shift in AI isn't just making pretraining bigger. It's figuring out how much capability can be created after pretraining.
Introducing Naive-N0.5-Flash: Building Frontier AI with AI
🔹 309B MoE, 15.5B active: top-tier in coding, leading in AI R&D.
🔹 Native 1M context, no full-attention layers (hybrid SWA + DSA).
🔹 Inference runtime built by AI: up to 2,000 tok/s in Ultrafast mode.
Weights are open today under MIT license.
🔗 Tech blog: naive.ai/en/research/
🤗 Hugging Face: huggingface.co/NaiveAI/Naive…
💻 GitHub: github.com/NaiveAI-Labs/Naiv…
🌐 naive.ai/en/
14
109
126
10,794
lee retweeted
Everyone will look at the benchmark numbers. I'm more curious about the process that produced them.
Introducing Naive-N0.5-Flash: Building Frontier AI with AI
🔹 309B MoE, 15.5B active: top-tier in coding, leading in AI R&D.
🔹 Native 1M context, no full-attention layers (hybrid SWA + DSA).
🔹 Inference runtime built by AI: up to 2,000 tok/s in Ultrafast mode.
Weights are open today under MIT license.
🔗 Tech blog: naive.ai/en/research/
🤗 Hugging Face: huggingface.co/NaiveAI/Naive…
💻 GitHub: github.com/NaiveAI-Labs/Naiv…
🌐 naive.ai/en/
14
108
120
10,093
lee retweeted
2,000+ tok/s as a peak single-request decode speed is a pretty wild engineering result. But I'm more interested in how they got there.
Introducing Naive-N0.5-Flash: Building Frontier AI with AI
🔹 309B MoE, 15.5B active: top-tier in coding, leading in AI R&D.
🔹 Native 1M context, no full-attention layers (hybrid SWA + DSA).
🔹 Inference runtime built by AI: up to 2,000 tok/s in Ultrafast mode.
Weights are open today under MIT license.
🔗 Tech blog: naive.ai/en/research/
🤗 Hugging Face: huggingface.co/NaiveAI/Naive…
💻 GitHub: github.com/NaiveAI-Labs/Naiv…
🌐 naive.ai/en/
13
109
155
17,121
lee retweeted
The interesting part isn't just the model. It's the idea of AI becoming part of the loop that builds the next generation of AI.
Introducing Naive-N0.5-Flash: Building Frontier AI with AI
🔹 309B MoE, 15.5B active: top-tier in coding, leading in AI R&D.
🔹 Native 1M context, no full-attention layers (hybrid SWA + DSA).
🔹 Inference runtime built by AI: up to 2,000 tok/s in Ultrafast mode.
Weights are open today under MIT license.
🔗 Tech blog: naive.ai/en/research/
🤗 Hugging Face: huggingface.co/NaiveAI/Naive…
💻 GitHub: github.com/NaiveAI-Labs/Naiv…
🌐 naive.ai/en/
14
108
121
10,613
lee retweeted
A new frontier lab shipping its first model with open weights makes this launch more interesting than a typical model announcement
Introducing Naive-N0.5-Flash: Building Frontier AI with AI
🔹 309B MoE, 15.5B active: top-tier in coding, leading in AI R&D.
🔹 Native 1M context, no full-attention layers (hybrid SWA + DSA).
🔹 Inference runtime built by AI: up to 2,000 tok/s in Ultrafast mode.
Weights are open today under MIT license.
🔗 Tech blog: naive.ai/en/research/
🤗 Hugging Face: huggingface.co/NaiveAI/Naive…
💻 GitHub: github.com/NaiveAI-Labs/Naiv…
🌐 naive.ai/en/
14
108
126
10,151
lee retweeted
Okay, the “AI helping build AI” part actually got my attention
Introducing Naive-N0.5-Flash: Building Frontier AI with AI
🔹 309B MoE, 15.5B active: top-tier in coding, leading in AI R&D.
🔹 Native 1M context, no full-attention layers (hybrid SWA + DSA).
🔹 Inference runtime built by AI: up to 2,000 tok/s in Ultrafast mode.
Weights are open today under MIT license.
🔗 Tech blog: naive.ai/en/research/
🤗 Hugging Face: huggingface.co/NaiveAI/Naive…
💻 GitHub: github.com/NaiveAI-Labs/Naiv…
🌐 naive.ai/en/
14
109
155
18,057
lee retweeted
Model launches usually tell you what was built. This one also gives a glimpse into how AI is changing the way models get built.
Introducing Naive-N0.5-Flash: Building Frontier AI with AI
🔹 309B MoE, 15.5B active: top-tier in coding, leading in AI R&D.
🔹 Native 1M context, no full-attention layers (hybrid SWA + DSA).
🔹 Inference runtime built by AI: up to 2,000 tok/s in Ultrafast mode.
Weights are open today under MIT license.
🔗 Tech blog: naive.ai/en/research/
🤗 Hugging Face: huggingface.co/NaiveAI/Naive…
💻 GitHub: github.com/NaiveAI-Labs/Naiv…
🌐 naive.ai/en/
14
110
156
20,910
lee retweeted
AI building AI is becoming less of a thought experiment
Introducing Naive-N0.5-Flash: Building Frontier AI with AI
🔹 309B MoE, 15.5B active: top-tier in coding, leading in AI R&D.
🔹 Native 1M context, no full-attention layers (hybrid SWA + DSA).
🔹 Inference runtime built by AI: up to 2,000 tok/s in Ultrafast mode.
Weights are open today under MIT license.
🔗 Tech blog: naive.ai/en/research/
🤗 Hugging Face: huggingface.co/NaiveAI/Naive…
💻 GitHub: github.com/NaiveAI-Labs/Naiv…
🌐 naive.ai/en/
14
107
122
10,755
lee retweeted
The more interesting benchmark here might be the research process itself. What happens when AI becomes part of the team building the next model?
Introducing Naive-N0.5-Flash: Building Frontier AI with AI
🔹 309B MoE, 15.5B active: top-tier in coding, leading in AI R&D.
🔹 Native 1M context, no full-attention layers (hybrid SWA + DSA).
🔹 Inference runtime built by AI: up to 2,000 tok/s in Ultrafast mode.
Weights are open today under MIT license.
🔗 Tech blog: naive.ai/en/research/
🤗 Hugging Face: huggingface.co/NaiveAI/Naive…
💻 GitHub: github.com/NaiveAI-Labs/Naiv…
🌐 naive.ai/en/
109
113
10,261