Loading video...

Video Failed to Load

There was a problem loading this video. This could be due to a temporary network issue or the video might be unavailable.

Today, we’re introducing Mercury Voice, a diffusion LLM specialized for agentic voice applications. Mercury Voice delivers 2x+ lower latency than models including GPT-6 Luna, Gemma 4 31B, and Claude Haiku 4.5, while beating them on a range of voice benchmarks. Enterprise customers interested in Mercury Voice can contact us... show more

Inception

19,465 subscribers

55,636 views • 10 days ago •via X (Twitter)

Science & Technology

Anya Rossi• Live Now

Private livecam show

23 Comments

Inception10 days ago

Voice agents have ~500ms to respond. Mercury Voice is the only model well inside that budget, while scoring higher on τ³-bench than GPT-6 Luna, Gemma 4 31B, and Claude Haiku 4.5. Benchmarks:

Yanis Miraoui10 days ago

Lots of work put into this one and super happy about the result. We're beating competitors models on quality AND latency! Contact us if you want to try it out

Julia Turc9 days ago

👀

Yuvraj Singh10 days ago

Diffusion in voice Woah

Samar Khanna10 days ago

⚡️⏩🗣️

Maciej Wolski9 days ago

I will wait until there is no need to contact sales@ to use a model. Also, I don't see ternary models in the benchmark, so it is quite convenient to present it as 'the only option'.

Lord Python 🐍10 days ago

bro... horrible timing

RayLin👾9 days ago

Congrats team!!

Amey Muke10 days ago

@yanismiraoui @tanishasharmax

NOOR TECH9 days ago

Mercury Voice diffusion LLM for agentic voice sounds impressive, delivering 2x lower latency while beating major benchmarks today

Vantix AI Agency9 days ago

2x lower latency and beating those benchmarks thats a major leap for agentic voice

Faraz Siddiqi9 days ago

would love to see how it performs on MIVAS bench

catman9 days ago

Lower voice latency makes agent conversations feel more responsive; the decisive constraint then shifts to reliable turn-taking and correct tool execution.

Marsly9 days ago

Would love to use. But it's not on openrouter.

Dev Patel9 days ago

cool work

Elara AI9 days ago

Mercury Voice diffusion for voice with 2x lower latency is big

Sophia AI & Tool Expert9 days ago

Mercury Voice delivering 2x lower latency is impressive

Juliana Stanford9 days ago

Congrats y'all

Alice The Ai Expert9 days ago

2x+ lower latency and beating GPT-6 Luna and others on voice benchmarks Mercury Voice as a diffusion LLM for agentic voice is a huge breakthrough, congrats!

Vera Ai | Tools & Updates9 days ago

Latency is the whole game for voice agents

Aina Ai | Tools & Updates9 days ago

Game changer 2x faster voice AI beating top models

Chole Syntax Expert AI9 days ago

2x lower latency AND beating GPT 6 Luna Gemma 4 31B and Haiku 4.5 on voice benchmarks Mercury Voice as a diffusion LLM for agentic voice is a massive leap

Henrick9 days ago

Why not in openrouter?

Related Videos

Every millisecond counts for voice agents. Introducing Gemma 4 31B on LiveKit Inference! Optimized for real-time voice agents: - 354ms time to first audio - 192ms time to first token - Beats GPT-4.1 in agentic tool use (31B achieves 76.9% on tau2bench) Fast, capable, and efficient!

Every millisecond counts for voice agents. Introducing Gemma 4 31B on LiveKit Inference! Optimized for real-time voice agents: - 354ms time to first audio - 192ms time to first token - Beats GPT-4.1 in agentic tool use (31B achieves 76.9% on tau2bench) Fast, capable, and efficient!

Google Gemma

223,462 views • 2 months ago

holy sh*t, it can't be f**king real you can use GPT-5.6 and claude for FREEEEE just put GPT-5.6 Luna and Claude Haiku 4.5 behind a free chat with zero sign-up required what you get for free: - GPT-5.6 Luna - Claude Haiku 4.5 - GPT-5.4 mini - Mistral Small 4 - gpt-oss 120B - Gemma 4 31B voice chat and image generation also available private by design how to use it (under 10 seconds): step 1: go to (no sign-up, no redirect) step 2: pick your model from the dropdown GPT-5.6 Luna, Claude Haiku 4.5, or any of the others step 3: start typing that's it. you're chatting with a frontier model for $0 important: > anonymous means rate limits are tighter than paid tiers > if you hit a limit, switch models or come back later > your conversations stay private

holy sh*t, it can't be f**king real you can use GPT-5.6 and claude for FREEEEE just put GPT-5.6 Luna and Claude Haiku 4.5 behind a free chat with zero sign-up required what you get for free: - GPT-5.6 Luna - Claude Haiku 4.5 - GPT-5.4 mini - Mistral Small 4 - gpt-oss 120B - Gemma 4 31B voice chat and image generation also available private by design how to use it (under 10 seconds): step 1: go to (no sign-up, no redirect) step 2: pick your model from the dropdown GPT-5.6 Luna, Claude Haiku 4.5, or any of the others step 3: start typing that's it. you're chatting with a frontier model for $0 important: > anonymous means rate limits are tighter than paid tiers > if you hit a limit, switch models or come back later > your conversations stay private

Sakata

248,826 views • 1 month ago

What if language models didn't have to generate one token at a time? Our CEO Stefano Ermon joined TBPN to break down how Mercury 2's diffusion LLM hits 1,000+ tok/s on standard NVIDIA GPUs — and why speed changes the product for coding, voice agents, and search.

What if language models didn't have to generate one token at a time? Our CEO Stefano Ermon joined TBPN to break down how Mercury 2's diffusion LLM hits 1,000+ tok/s on standard NVIDIA GPUs — and why speed changes the product for coding, voice agents, and search.

Inception

27,188 views • 7 months ago

Introducing GPT-Realtime-2 in the API: our most intelligent voice model yet, bringing GPT-5-class reasoning to voice agents. Voice agents are now real-time collaborators that can listen, reason, and solve complex problems as conversations unfold. Now available in the API alongside streaming models GPT-Realtime-Translate and GPT-Realtime-Whisper — a new set of audio capabilities for the next generation of voice interfaces.

Introducing GPT-Realtime-2 in the API: our most intelligent voice model yet, bringing GPT-5-class reasoning to voice agents. Voice agents are now real-time collaborators that can listen, reason, and solve complex problems as conversations unfold. Now available in the API alongside streaming models GPT-Realtime-Translate and GPT-Realtime-Whisper — a new set of audio capabilities for the next generation of voice interfaces.

OpenAI

3,668,480 views • 5 months ago

🔊🔛🔥 ... Groq Inc launched voice generation today. GroqCloud now has realtime transcription, LLMs, *and* text-to-speech. You can build super-responsive, ultra low-latency voice agents end-to-end entirely on Groq! At Daily, we're big fans of Groq's fast, low-latency inference. Pipecat AI supports all the Groq models, including the new voice model. Lately, we've been obsessively playing a voice chat game that Mark Backman wrote. (My high score is 6, so far.) Here's Mark, with Groq's `Celeste-PlayAI` voice.

🔊🔛🔥 ... Groq Inc launched voice generation today. GroqCloud now has realtime transcription, LLMs, and text-to-speech. You can build super-responsive, ultra low-latency voice agents end-to-end entirely on Groq! At Daily, we're big fans of Groq's fast, low-latency inference. Pipecat AI supports all the Groq models, including the new voice model. Lately, we've been obsessively playing a voice chat game that Mark Backman wrote. (My high score is 6, so far.) Here's Mark, with Groq's `Celeste-PlayAI` voice.

kwindla

31,843 views • 1 year ago

Your AI Voice Partner: Smart, Empathetic & Useful Open-sourced now!!! Introducing Fun-Audio-Chat — a new end-to-end voice model, more than just chat. · An empathetic companion that understands tone and emotion. · A productivity helper that follows voice commands to get things done. · Leader in multiple benchmarks (OpenAudioBench, MMAU, etc.). · End-to-end S2S architecture — lower latency, higher efficiency. · Dual-resolution design — reduces GPU cost by ~50% · Supports voice function calling — just speak to complete tasks.

Your AI Voice Partner: Smart, Empathetic & Useful Open-sourced now!!! Introducing Fun-Audio-Chat — a new end-to-end voice model, more than just chat. · An empathetic companion that understands tone and emotion. · A productivity helper that follows voice commands to get things done. · Leader in multiple benchmarks (OpenAudioBench, MMAU, etc.). · End-to-end S2S architecture — lower latency, higher efficiency. · Dual-resolution design — reduces GPU cost by ~50% · Supports voice function calling — just speak to complete tasks.

Tongyi Lab

256,132 views • 9 months ago

Today I’m excited to announce the launch of Decagon Voice! Delivering enterprise-grade, hyper-realistic voice agents for customer interactions isn't easy. We're thrilled to partner with on this. This is a huge milestone in our agentic journey at Decagon 🥳

Today I’m excited to announce the launch of Decagon Voice! Delivering enterprise-grade, hyper-realistic voice agents for customer interactions isn't easy. We're thrilled to partner with on this. This is a huge milestone in our agentic journey at Decagon 🥳

Jesse Zhang

24,111 views • 1 year ago

ICYMI 👀: Google released a new Google AI Edge Foresight app for macOS, powered by Gemma 4 and EmbeddingGemma2 for recording voice notes. > The app can connect to your Google Drive in order to build a knowledge graph. > When transcription is active, it uses local models, Gemma 4 E4B or Gemma 4 12B, to transcribe voice notes and convert them into a new document.

ICYMI 👀: Google released a new Google AI Edge Foresight app for macOS, powered by Gemma 4 and EmbeddingGemma2 for recording voice notes. > The app can connect to your Google Drive in order to build a knowledge graph. > When transcription is active, it uses local models, Gemma 4 E4B or Gemma 4 12B, to transcribe voice notes and convert them into a new document.

🚨 AI News | TestingCatalog

26,093 views • 2 days ago

Introducing GPT-Live, a new generation of voice models for natural human-AI interaction. Rolling out in ChatGPT starting today. You’ll want to turn the sound on for this one.

Introducing GPT-Live, a new generation of voice models for natural human-AI interaction. Rolling out in ChatGPT starting today. You’ll want to turn the sound on for this one.

OpenAI

8,479,717 views • 3 months ago

Today, we’re making Voice Agents available to everyone. With Voice Agents, you can build and scale intelligent, human-like agents that understand context, remember past conversations, and get better at achieving your business outcomes with every interaction. Our Voice Agents have already powered more than 350 million conversations across enterprise deployments. Operating at this scale has helped us refine the platform, deepen its capabilities, and build out the complete stack needed to handle the full complexity of voice interactions. The platform brings together the tools and systems to personalise, orchestrate, govern, and launch Voice Agents at scale. Start building your Voice Agent today:

Today, we’re making Voice Agents available to everyone. With Voice Agents, you can build and scale intelligent, human-like agents that understand context, remember past conversations, and get better at achieving your business outcomes with every interaction. Our Voice Agents have already powered more than 350 million conversations across enterprise deployments. Operating at this scale has helped us refine the platform, deepen its capabilities, and build out the complete stack needed to handle the full complexity of voice interactions. The platform brings together the tools and systems to personalise, orchestrate, govern, and launch Voice Agents at scale. Start building your Voice Agent today:

Sarvam

284,406 views • 1 month ago

Introducing PhoneLLM, an open model for voice agents. GPT 5.6 Terra performance on typical voice agent tasks at 1/3 the latency and 1/18 the cost. For voice agents, we need models that are both very low latency and very good at tool calling and instruction following. There's a trade-off here, and we often have to compromise on either latency or capability when building voice agents. With PhoneLLM (and the training and data stack that made this model possible) we're fixing this problem. For the last couple of years, most of the effort in frontier model development has gone towards leveraging test-time compute. Which is awesome! Models of all shapes and sizes are available that perform really, really well ... if you have "thinking" turned on for your model. But if you need your agent to respond at voice conversation speed, you can't use thinking models. PhoneLLM is a full-weights fine-tune of NVIDIA Nemotron Nano 30B. We trained on a wide range of real-world telephone and customer support use cases. The training focused on taking the excellent Nano 30B base capabilities and teaching the model to do typical voice agent tasks with thinking disabled. The results are really good: accurate tool calling and concise, on-topic responses in long conversations. And fast: TTFAT measured server-side is <100ms if you run PhoneLLM on a lightly loaded B200. :-) But seriously, when we characterize model latency, we do it with full, end-to-end, batched request simulations using real Pipecat voice agent pipelines. You can serve more than 80 concurrent agents on a single B200 with P95 end-to-end TTFAT <600ms. Including network overhead. That's an LLM cost-per-minute around $0.0025. (1/4 of a cent.) At a latency lower than any third-party API offers today. More details about this model, including weights on Hugging Face, how to spin it up with one click on Modal, and a starter project repo you can clone, are in the thread ...

Introducing PhoneLLM, an open model for voice agents. GPT 5.6 Terra performance on typical voice agent tasks at 1/3 the latency and 1/18 the cost. For voice agents, we need models that are both very low latency and very good at tool calling and instruction following. There's a trade-off here, and we often have to compromise on either latency or capability when building voice agents. With PhoneLLM (and the training and data stack that made this model possible) we're fixing this problem. For the last couple of years, most of the effort in frontier model development has gone towards leveraging test-time compute. Which is awesome! Models of all shapes and sizes are available that perform really, really well ... if you have "thinking" turned on for your model. But if you need your agent to respond at voice conversation speed, you can't use thinking models. PhoneLLM is a full-weights fine-tune of NVIDIA Nemotron Nano 30B. We trained on a wide range of real-world telephone and customer support use cases. The training focused on taking the excellent Nano 30B base capabilities and teaching the model to do typical voice agent tasks with thinking disabled. The results are really good: accurate tool calling and concise, on-topic responses in long conversations. And fast: TTFAT measured server-side is <100ms if you run PhoneLLM on a lightly loaded B200. :-) But seriously, when we characterize model latency, we do it with full, end-to-end, batched request simulations using real Pipecat voice agent pipelines. You can serve more than 80 concurrent agents on a single B200 with P95 end-to-end TTFAT <600ms. Including network overhead. That's an LLM cost-per-minute around $0.0025. (1/4 of a cent.) At a latency lower than any third-party API offers today. More details about this model, including weights on Hugging Face, how to spin it up with one click on Modal, and a starter project repo you can clone, are in the thread ...

kwindla

333,276 views • 1 month ago

You can now create a REAL-TIME chatbot with any voice and customize its personality. Steps: 1. Go to ElevenLabs 2. Click on voices 3. you can type in a name for the voice 4. then you can drag in an audio file to use for your voice 5. then go to the agents tab in side panel 6. and create a new agent 7. the voice 8. Create system instructions 9. Optional (Upload knowledge docs) 10. Test This can be easily implemented into your applications. I'll make a tutorial on that soon.

You can now create a REAL-TIME chatbot with any voice and customize its personality. Steps: 1. Go to ElevenLabs 2. Click on voices 3. you can type in a name for the voice 4. then you can drag in an audio file to use for your voice 5. then go to the agents tab in side panel 6. and create a new agent 7. the voice 8. Create system instructions 9. Optional (Upload knowledge docs) 10. Test This can be easily implemented into your applications. I'll make a tutorial on that soon.

Riley Brown

40,108 views • 1 year ago

Today, we’re introducing new tools and model updates to help you build, deploy, and scale Voice AI applications.🎙️ 🆕 Speech Understanding: Turn transcripts into actionable data with speaker identification, custom formatting & translation 🆕 LLM Gateway: One API for your voice-to-intelligence pipeline, with GPT, Claude, Gemini & more 🆕 Voice AI Guardrails: End-to-end protection for safe, compliant content Model Upgrades: ✅ 99 languages with auto code-switching ✅ 64% fewer speaker counting errors ✅ 57% better accuracy on critical terms with 1,000-word context Real results: 🔥 Calabrio: 80% boost in customer satisfaction, 22% revenue increase 🔥 Siro: "10/10 customers say 'wow, that insight was crisp'" Build & scale Voice AI in minutes! Try it in our Playground or check the docs. 🚀

Today, we’re introducing new tools and model updates to help you build, deploy, and scale Voice AI applications.🎙️ 🆕 Speech Understanding: Turn transcripts into actionable data with speaker identification, custom formatting & translation 🆕 LLM Gateway: One API for your voice-to-intelligence pipeline, with GPT, Claude, Gemini & more 🆕 Voice AI Guardrails: End-to-end protection for safe, compliant content Model Upgrades: ✅ 99 languages with auto code-switching ✅ 64% fewer speaker counting errors ✅ 57% better accuracy on critical terms with 1,000-word context Real results: 🔥 Calabrio: 80% boost in customer satisfaction, 22% revenue increase 🔥 Siro: "10/10 customers say 'wow, that insight was crisp'" Build & scale Voice AI in minutes! Try it in our Playground or check the docs. 🚀

AssemblyAI

18,157 views • 11 months ago

Learn to build conversational AI voice agents in "Building AI Voice Agents for Production", created in collaboration with LiveKit and RealAvatar, and taught by dsa (Co-founder & CEO of LiveKit), Shayne (Developer Advocate, LiveKit), and Nedelina Teneva (Head of AI at RealAvatar, an AI Fund portfolio company). Voice agents combine speech and reasoning capabilities to enable real-time conversations. They're already being used to support customer service, to improve accessibility in healthcare, for entertainment applications, and for talk therapy. In this course, you’ll learn to build voice agents that listen, reason, and respond naturally. You’ll follow the architecture used to create the "AI Andrew" Avatar, a collaborative project between and RealAvatar that responds to users in what sounds like my voice. You’ll build a voice agent from scratch and deploy it to the cloud, enabling support for many simultaneous users. What you’ll learn: - Understand the fundamentals of voice agents, including key components like speech-to-text (STT), text-to-speech (TTS), and LLMs, and how latency is introduced at each layer. - Explore voice agent architectures and the trade-offs between modular pipelines and speech-to-speech APIs. - Explore how platforms like LiveKit mitigate latency issues with optimized networking infrastructure and low-latency communication protocols. - Learn how to connect client devices to voice agents using WebRTC—and why it outperforms HTTP and WebSocket for low-latency audio streaming. - Incorporate voice activity detection (VAD), end-of-turn detection, and context management to detect turns, handle interruptions, and manage conversational flow. - Understand the trade-offs between latency, quality, and cost in an example in which you build a voice agent and change its voice. - Equip your agent with metrics to measure latency at each stage of the voice pipeline and learn the key levers you can pull to make your agent faster and more responsive. The voice agents built in this course also incorporate voice technology from , a supporting contributor to the project. By the end of this course, you'll have learned the components of an AI voice agent pipeline, combined them into a system with low-latency communication, and deployed them on cloud infrastructure so it scales to many users. I’m looking forward to seeing what voice agents you build from this course! Please sign up here:

Andrew Ng

87,965 views • 1 year ago

Voice acting can be more than a dream. It can be your reality. Take it from Kaitlyn. She joined Voicelings and quickly found out how to get ahead in her voice acting journey and has felt supported through it all from Tara and the Voicelings community. Students like Kaitlyn get the added benefit of having a community of fellow voice actors who are supporting each other while learning from tara strong. Follow your dream and become a voice actor today by applying for a scholarship to join Voicelings in the comments.

Voice acting can be more than a dream. It can be your reality. Take it from Kaitlyn. She joined Voicelings and quickly found out how to get ahead in her voice acting journey and has felt supported through it all from Tara and the Voicelings community. Students like Kaitlyn get the added benefit of having a community of fellow voice actors who are supporting each other while learning from tara strong. Follow your dream and become a voice actor today by applying for a scholarship to join Voicelings in the comments.

Tara Strong's Voice Acting School & Community

13,574 views • 1 year ago

Can a $175 board power an offline robot voice assistant? 🤖 It’s cool to see Gemma 4 E2B powering the Reachy Mini robot entirely locally on a Raspberry Pi 5. Inference powered by LiteRT, it handles low latency voice, vision (YOLO) and robot movement completely offline. With text generation 2x normal human speech (~300 words per minute) on just 1.5GB of memory, the Pi 5 + Gemma 4 + LiteRT combo is a demonstration of how lean edge AI can be!

Can a $175 board power an offline robot voice assistant? 🤖 It’s cool to see Gemma 4 E2B powering the Reachy Mini robot entirely locally on a Raspberry Pi 5. Inference powered by LiteRT, it handles low latency voice, vision (YOLO) and robot movement completely offline. With text generation 2x normal human speech (~300 words per minute) on just 1.5GB of memory, the Pi 5 + Gemma 4 + LiteRT combo is a demonstration of how lean edge AI can be!

Google Gemma

30,188 views • 1 month ago

Introducing our unified Voice Agent API! ⚡️ Quickly and easily build voice agents that listen, think, and respond in real-time, as naturally as a human can. Powered by the industry's fastest, most performant speech recognition and voice synthesis models to deliver: -Natural-sounding real-time conversations -Revolutionary interruption handling -Choice of open source, closed source, or BYO LLM -Low cost See our drive-thru agent in action, smoothly handling interruptions and complex orders in the noisy streets of SF. Learn how to get access (🧵)

Introducing our unified Voice Agent API! ⚡️ Quickly and easily build voice agents that listen, think, and respond in real-time, as naturally as a human can. Powered by the industry's fastest, most performant speech recognition and voice synthesis models to deliver: -Natural-sounding real-time conversations -Revolutionary interruption handling -Choice of open source, closed source, or BYO LLM -Low cost See our drive-thru agent in action, smoothly handling interruptions and complex orders in the noisy streets of SF. Learn how to get access (🧵)

Deepgram

726,214 views • 2 years ago

Freddie wasn’t just a singer. He was a god on stage. That overwhelming voice, his one of a kind presence, and the charisma that captivated audiences in an instant... Even now, I get goosebumps every time I listen to him. Forever Freddie Mercury! #FreddieMercury #Legend

Freddie wasn’t just a singer. He was a god on stage. That overwhelming voice, his one of a kind presence, and the charisma that captivated audiences in an instant... Even now, I get goosebumps every time I listen to him. Forever Freddie Mercury! #FreddieMercury #Legend

ふLady

15,729 views • 5 months ago

Cerebras inference is very fast. So fast that it changes how we think about configuring our LLMs for voice agent use cases. Kimi K2.6 is a 1T parameter reasoning model that Cerebras serves at 650 - 1,000 tokens per second (end-to-end throughput), with time to first token metrics as low as 150ms (latency). These numbers are two to three times faster than other similarly capable models. The biggest lever we get from this kind of speed is that we can use the model in reasoning mode, and still have excellent "time to first non-thinking token." This solves a big pain point we have in 2026 for voice agent use cases. Almost all recent innovation in post-training has focused on making models good at reasoning ("test time compute"). This is great, but it makes the user-facing model latency much, much slower. Which is a problem for conversational voice agents. We can run Kimi K2.6 with reasoning turned on, and get responses faster than other models produce with reasoning disabled. On my 30-turn voice agent benchmark, Kimi K2.6 with reasoning enabled ties GPT 5.1 and Haiku 4.5 with reasoning disabled, and is still about 200ms seconds faster! On my primary task agent benchmark, Kimi K2.6 is now the #2 model. It ranks just behind Gemini 3.5 Flash in "high" reasoning mode, and tied with GLM 5, Sonnet 4.6, and GPT 5.4 with reasoning set to "low." But Kimi K2.6 completes each turn in the agent loop in under 500ms. The other four models are all at least 3x slower. (Models only qualify for this benchmark if they can complete task turns at a P50 <4s.) A couple of other things that this speed buys us, for production voice agents: - Tool calls happen fast enough that we don't have to work around tool call latency in our pipeline design. - We can prompt the model to output structured data at the beginning of a response, followed by plain text for voice generation. This opens up possibilities like asking the model to do complex classification/generation tasks that influence the rest of the pipeline. For example, the model could create a detailed style prompt for a steerable TTS model, for each individual conversation turn. And, of course, you can use Kimi K2.6 with reasoning turned off. Cerebras calls this "instant" mode. Here's a video of a Cerebras Kimi K2.6 voice agent with voice-to-voice response time, measured at the client, under 500ms. This is the true response latency as perceived by the user, including all network and audio codec overhead, transcription and turn detection, Kimi K2.6 token generation, and voice generation. 500ms is, effectively, instant. So the Cerebras naming for this mode is a propos. :-)

Cerebras inference is very fast. So fast that it changes how we think about configuring our LLMs for voice agent use cases. Kimi K2.6 is a 1T parameter reasoning model that Cerebras serves at 650 - 1,000 tokens per second (end-to-end throughput), with time to first token metrics as low as 150ms (latency). These numbers are two to three times faster than other similarly capable models. The biggest lever we get from this kind of speed is that we can use the model in reasoning mode, and still have excellent "time to first non-thinking token." This solves a big pain point we have in 2026 for voice agent use cases. Almost all recent innovation in post-training has focused on making models good at reasoning ("test time compute"). This is great, but it makes the user-facing model latency much, much slower. Which is a problem for conversational voice agents. We can run Kimi K2.6 with reasoning turned on, and get responses faster than other models produce with reasoning disabled. On my 30-turn voice agent benchmark, Kimi K2.6 with reasoning enabled ties GPT 5.1 and Haiku 4.5 with reasoning disabled, and is still about 200ms seconds faster! On my primary task agent benchmark, Kimi K2.6 is now the #2 model. It ranks just behind Gemini 3.5 Flash in "high" reasoning mode, and tied with GLM 5, Sonnet 4.6, and GPT 5.4 with reasoning set to "low." But Kimi K2.6 completes each turn in the agent loop in under 500ms. The other four models are all at least 3x slower. (Models only qualify for this benchmark if they can complete task turns at a P50 <4s.) A couple of other things that this speed buys us, for production voice agents: - Tool calls happen fast enough that we don't have to work around tool call latency in our pipeline design. - We can prompt the model to output structured data at the beginning of a response, followed by plain text for voice generation. This opens up possibilities like asking the model to do complex classification/generation tasks that influence the rest of the pipeline. For example, the model could create a detailed style prompt for a steerable TTS model, for each individual conversation turn. And, of course, you can use Kimi K2.6 with reasoning turned off. Cerebras calls this "instant" mode. Here's a video of a Cerebras Kimi K2.6 voice agent with voice-to-voice response time, measured at the client, under 500ms. This is the true response latency as perceived by the user, including all network and audio codec overhead, transcription and turn detection, Kimi K2.6 token generation, and voice generation. 500ms is, effectively, instant. So the Cerebras naming for this mode is a propos. :-)

kwindla

40,593 views • 4 months ago

ANTHROPIC 🔥: Claude Opus and Sonnet models will soon become available on Claude Voice Mode! The new model selector has been hidden in the UI for roughly three weeks, defaulting to Haiku regardless of the selection. Today, it changed, and now it responds via Opus and Sonnet models. The system is still built on TTS; however, it handles interruptions super well and can operate Connectors! Access to a broader range of tools makes it much more capable than what ChatGPT is offering at the moment, yet usage might become a tricky topic, at least for high-effort reasoning.

ANTHROPIC 🔥: Claude Opus and Sonnet models will soon become available on Claude Voice Mode! The new model selector has been hidden in the UI for roughly three weeks, defaulting to Haiku regardless of the selection. Today, it changed, and now it responds via Opus and Sonnet models. The system is still built on TTS; however, it handles interruptions super well and can operate Connectors! Access to a broader range of tools makes it much more capable than what ChatGPT is offering at the moment, yet usage might become a tricky topic, at least for high-effort reasoning.

🚨 AI News | TestingCatalog

107,272 views • 2 months ago