Mercury Voice is a diffusion-based language model built by Inception Labs for teams shipping real-time voice agents, the kind of software that has to reason, call tools, and follow long system prompts without leaving a caller in silence. The launch marks general availability for enterprise customers after a preview two weeks ago alongside Mercury 2.5, and it targets a specific pain point that has forced voice builders to pick between fast non-reasoning models and slower, smarter ones.
On Inception's internal test set of production voice-agent prompts, Mercury Voice returns its first answer token in a median of 320 milliseconds, keeping it inside the roughly 500 ms window that makes a spoken exchange feel natural. The company reports it beats GPT-6 Luna, Gemma 4 31B, and GLM-5.3-Flash on a composite that includes τ³-bench Telecom, Retail, and Airline, IFBench, and BFCL v4 , and it ships with three reasoning-effort settings, a 128K context window with up to 50K output tokens, and launch pricing of $0.20 per million input and $0.75 per million output tokens (half of the standard $0.40 and $1.50) . One early production user, Audivi AI, is running it for fully automated drive-thru ordering.
Inception Labs was founded by Stanford professor Stefano Ermon, who co-invented diffusion models, along with UCLA professor Aditya Grover and Cornell professor Volodymyr Kuleshov . The company applies the diffusion approach that underpins image and video generators to text, and recently closed a $50M round led by Menlo Ventures with participation from Andrew Ng and Andrej Karpathy , positioning Mercury Voice as its most latency-sensitive product to date.
Comments (12)
Sign in to join the discussion.
T
Tomas Lindqvist14d ago
Naming your competitor 'GPT-6 Luna' in your launch tweet is a bold way to manifest a lawsuit before lunch.
P
Priyanshu R.14d ago
Diffusion for voice latency actually makes sense, but I want to see the p99 under real load, not a curated clip. Is the demo live or was it rendered at 3am by a sleepy intern?
S
Sofia Valdez14d ago
Been tracking diffusion LLM teams since the first Mercury drop and this is the arc I wanted to see. Voice was the obvious wedge.
M
Mina Haddad14d ago
The tweet buries the lede. 'Every millisecond matters' should be the first four words, not the tagline hiding behind three benchmark names.
K
Kwesi Otieno14d ago
Rewrite: 'Voice agents that answer before you finish asking.' You can venmo me later.
Y
yuki14d ago
Where are the docs? If I can't curl a hello-world in under 4 minutes I'm going back to whatever I was procrastinating on.
B
Burak E.14d ago
2x lower latency is great until the pricing page loads. Per-minute? Per-token? Per-vibe? Tell me how this hits gross margin before I let my team ship it.
C
Chidinma Okafor14d ago
Voice agents are cool but has anyone considered that in 2019 we called this IVR and everyone hated it. Curious what changed besides the price of GPUs.
R
Ravi Teja14d ago
First 3 seconds of that demo audio are pure gold, the rest drags. If you're pitching latency, cut the video in half and lead with the interrupt handling.
H
Hallie Engreen14d ago
Speed is table stakes now, the moat is interruption grace. Does Mercury Voice handle overlapping speech or does it panic like the rest of them?
N
Noor Ben Ali14d ago
Any chance a slim variant of this ends up open weights, even research-only? DevRel plea from someone who has to demo without an enterprise contract.
D
Dmitri Koval14d ago
Great, another model I'll have to explain to my mom is not the same as the token she bought last Tuesday.
Work With Us
Book an intro call to explore our suite of services. We might have a waitlist.
Inception Labs
Every Millisecond Matters
Every Millisecond Matters
Tags
TLVC Rating
Community Rating
No ratings yet
to rate this launch.
Mercury Voice is a diffusion-based language model built by Inception Labs for teams shipping real-time voice agents, the kind of software that has to reason, call tools, and follow long system prompts without leaving a caller in silence. The launch marks general availability for enterprise customers after a preview two weeks ago alongside Mercury 2.5, and it targets a specific pain point that has forced voice builders to pick between fast non-reasoning models and slower, smarter ones. On Inception's internal test set of production voice-agent prompts, Mercury Voice returns its first answer token in a median of 320 milliseconds, keeping it inside the roughly 500 ms window that makes a spoken exchange feel natural. The company reports it beats GPT-6 Luna, Gemma 4 31B, and GLM-5.3-Flash on a composite that includes τ³-bench Telecom, Retail, and Airline, IFBench, and BFCL v4 , and it ships with three reasoning-effort settings, a 128K context window with up to 50K output tokens, and launch pricing of $0.20 per million input and $0.75 per million output tokens (half of the standard $0.40 and $1.50) . One early production user, Audivi AI, is running it for fully automated drive-thru ordering. Inception Labs was founded by Stanford professor Stefano Ermon, who co-invented diffusion models, along with UCLA professor Aditya Grover and Cornell professor Volodymyr Kuleshov . The company applies the diffusion approach that underpins image and video generators to text, and recently closed a $50M round led by Menlo Ventures with participation from Andrew Ng and Andrej Karpathy , positioning Mercury Voice as its most latency-sensitive product to date.
Comments (12)
Naming your competitor 'GPT-6 Luna' in your launch tweet is a bold way to manifest a lawsuit before lunch.
Diffusion for voice latency actually makes sense, but I want to see the p99 under real load, not a curated clip. Is the demo live or was it rendered at 3am by a sleepy intern?
Been tracking diffusion LLM teams since the first Mercury drop and this is the arc I wanted to see. Voice was the obvious wedge.
The tweet buries the lede. 'Every millisecond matters' should be the first four words, not the tagline hiding behind three benchmark names.
Rewrite: 'Voice agents that answer before you finish asking.' You can venmo me later.
Where are the docs? If I can't curl a hello-world in under 4 minutes I'm going back to whatever I was procrastinating on.
2x lower latency is great until the pricing page loads. Per-minute? Per-token? Per-vibe? Tell me how this hits gross margin before I let my team ship it.
Voice agents are cool but has anyone considered that in 2019 we called this IVR and everyone hated it. Curious what changed besides the price of GPUs.
First 3 seconds of that demo audio are pure gold, the rest drags. If you're pitching latency, cut the video in half and lead with the interrupt handling.
Speed is table stakes now, the moat is interruption grace. Does Mercury Voice handle overlapping speech or does it panic like the rest of them?
Any chance a slim variant of this ends up open weights, even research-only? DevRel plea from someone who has to demo without an enterprise contract.
Great, another model I'll have to explain to my mom is not the same as the token she bought last Tuesday.