
Lucas Bunzel retweeted
Yesterday, we launched Mercury 3, our fastest model to date with speeds over 1000+ tokens/second on @nvidia hardware.
In this demo from @stefan_webb, we compare Mercury 3 + @opencode against Claude Code with Sonnet 5.5 and other model/harness combinations on a coding task utilizing many subagents running in parallel. (🧵1/2)
Lucas Bunzel retweeted
Could diffusion LLMs upend the AI industry?
I talked to @_inception_ai CEO @StefanoErmon about Mercury 3, which delivers faster, cheaper AI inference using a radically different architecture.
The implications go well beyond speed.
thedeepview.com/articles/why…
Lucas Bunzel retweeted
📰 Just in, @_inception_ai's diffusion-based LLM, Mercury 2, has been recognized by TIME as a best invention of 2026 in AI
Beside such distinguished company as @claudeai Claude Code, @NVIDIAAI NemoClaw, and @Lovable
🤫 More to follow next week...
Excited to share that Mercury 2 is one of @TIME's Best Inventions of 2026.
Not long ago, diffusion LLMs were just a research idea. Now Mercury powers products people use every day.
Proud of the @_inception_ai team. Something bigger coming next week 👇🧵
Lucas Bunzel retweeted
I had fun building this demo showing off the new vision capabilities of Mercury Decide!
Adding vision understanding to your decisions is as simple as adding a parameter to the query
👁️ Mercury Decide now accepts image input, which opens it up to a much wider range of applications:
openrouter.ai/inception/merc…
In the demo from @stefan_webb, Mercury Decide plays Ring King on NES, using Mercury Decide’s new vision capabilities to understand the state of the game and suggest the next action to take.
What about the other applications? (🧵 1/4)
Lucas Bunzel retweeted
👁️ Mercury Decide now accepts image input, which opens it up to a much wider range of applications:
openrouter.ai/inception/merc…
In the demo from @stefan_webb, Mercury Decide plays Ring King on NES, using Mercury Decide’s new vision capabilities to understand the state of the game and suggest the next action to take.
What about the other applications? (🧵 1/4)
Lucas Bunzel retweeted
Last week, we launched Mercury Decide, a Jev-beating decision model that quickly rose to #2 on OpenRouter by number of users: openrouter.ai/inception/merc…
This week, we’re updating Mercury Decide with substantially better benchmark performance, longer context (33k → 66k), General Availability with Zero Data Retention (ZDR), and more.
All that with lower pricing than Jev (ZDR for $0.02/$0 per 1M in/out, discounted at 50% from the list price of $0.04/$0, plus 100M free tokens), alongside an unlimited free tier.
What have we improved? Read on to find out! [🧵1/4]
Lucas Bunzel retweeted
One week ago we launched Mercury Decide, @_inception_ai System One decision model. It hit #2 on OpenRouter by users, and we've been improving it every day since.
Independent results from @ValsAI show that Mercury Decide is more accurate and more affordable than Jev.
Proud of how fast this team moves.
Free on @OpenRouter and available through the Inception API and @baseten. I'd love to see what you build.
openrouter.ai/inception/merc…
Lucas Bunzel retweeted
I've added more decision models to BulletBench.
They're insanely fast but still just a little bit worse at the game than LLMs.
The best one seems to be Mercury Decide.
It's about time Mercury ranks well on here, I pretty much built BulletBench originally after testing Mercury 2 (but it ended up ranking too low)
springprompt.com/benchmarks/…
Lucas Bunzel retweeted
I've been waiting for OpenAI's Decision API as a Jev competitor.
Interestingly, it scores behind Mercury, Jev and Clef!
Lucas Bunzel retweeted
🤖 Last week, I compiled an awesome list of Awesome Jev lists and found the most widespread applications for decision models were to be found in AI coding harnesses like Claude Code and OpenCode.
For example, one challenging task for coding agents is to detect Observation Prompt Injection (OPI). I compared a handful of decision models to see how they stack up, results in the table.
What is Observation Prompt Injection (OPI)? Let's dig a bit deeper! (🧵1/4)
Lucas Bunzel retweeted
This week we launched Mercury Voice, a dLLM optimized for voice agents with 2x+ lower latency than models like GPT-6 Luna.
In this dental clinic receptionist demo, median end-to-end latency stays under 300ms, tool calls included.
See (or hear?) it yourself in the model playground: chat.inceptionlabs.ai/voice
Enterprise customers interested in Mercury Voice can contact us at [email protected] to get access.
Lucas Bunzel retweeted
Introducing Mercury Decide: the most intelligent decision model on @OpenRouter (JevBench v1.4) and one of the fastest.
Up to 14 decisions per second. Watch it go head-to-head with Jev in chess.
Free for early access on OpenRouter: openrouter.ai/inception/merc…
Lucas Bunzel retweeted
Today's LLMs can't generate token 10 until token 9 exists.
@_inception_ai CEO @StefanoErmon thinks he's found the winner: diffusion LLMs that generate tokens in parallel.
New episode of No Priors, out now.
Lucas Bunzel retweeted
The right way to measure AI value: intelligence per dollar.
Our co-founder and CEO @StefanoErmon joined @YahooFinance Live to talk about Mercury 2.5, our newest and most capable diffusion LLM model yet, and why he thinks about it this way. Thanks for having him @BrookeDiPalma!
Lucas Bunzel retweeted
Mercury powers real-time voice systems with the tightest latency budgets.
@OpenCall_AI runs AI phone agents on live patient calls.
After moving off a specialized chip provider, they scaled from 300M to over 1B tokens per day, with median model response latency under 200ms.
Lucas Bunzel retweeted
Mercury 2.5 from @_inception_ai is now generally available on OpenRouter.
It's the fastest model on our Fastest Models leaderboard that doesn't require specialized hardware.
A diffusion LLM: it decodes tokens in parallel rather than one at a time. The gain is largest on code-heavy prompts.
Use it here: openrouter.ai/inception/merc…
Lucas Bunzel retweeted
most agent calls aren't the one writing the answer -- they're rewrites, reranks, compactions, and tool searches. that's where latency and cost compound. enter mercury 2.5: fast + cheap + smart enough to own that entire layer.
two sleeper previews in the launch post: dLLM optimized for voice agents + dLLM as a router to the best models!
congrats @StefanoErmon @volokuleshov @adityagrover_ and team @_inception_ai!
Today we're excited to announce Mercury 2.5
It’s the most capable diffusion LLM on the market. It is a 40% jump in intelligence over Mercury 2 and runs at over 1,100 tokens/sec on widely available @NVIDIAAI GPUs.
inceptionlabs.ai/blog/introd…
Lucas Bunzel retweeted
Mercury powers real-time systems with the tightest latency budgets.
@OpenCall_AI uses Mercury to run voice agents on live patient calls. After switching from an AI inference chip provider, p50 latency fell below 200 ms and p99 fell from several minutes to one second. This keeps multi-step reasoning inside the latency budget of a live call.
Lucas Bunzel retweeted
Today we're excited to announce Mercury 2.5
It’s the most capable diffusion LLM on the market. It is a 40% jump in intelligence over Mercury 2 and runs at over 1,100 tokens/sec on widely available @NVIDIAAI GPUs.
inceptionlabs.ai/blog/introd…
















