Inception Labs

Inception Labs is an American artificial intelligence research company founded in 2024 by Stanford professor Stefano Ermon, with co-founders Aditya Grover (UCLA) and Volodymyr Kuleshov (Cornell). The company is headquartered in Palo Alto and develops the Mercury family of diffusion-based large-language models, which generate text and code in parallel rather than through sequential token prediction. As of November 2025, Inception had raised $50 million in seed funding led by Menlo Ventures, with Mercury 2 launched in February 2026 and reported throughput exceeding 1,000 tokens per second.

At a glance

• Founded: 2024 in Palo Alto, California.
• Status: Private. Reported valuation approximately $500 million.
• Funding: $50 million seed round announced November 2025. Led by Menlo Ventures, with Mayfield, Innovation Endeavors, Microsoft M12, Snowflake Ventures, Databricks Ventures, and Nvidia NVentures participating. Angel investors include Andrew Ng and Andrej Karpathy.
• CEO: Stefano Ermon (co-founder, Stanford professor of computer science).
• Other notable leadership: Aditya Grover (co-founder, UCLA professor), Volodymyr Kuleshov (co-founder, Cornell professor).
• Open weights: None as of April 2026. Mercury models are accessed via API and partner integrations.
• Flagship models: Mercury (February 2025) and Mercury 2 (February 2026), available in general-purpose and Coder variants.

Origins

Inception Labs was founded in 2024 by Stefano Ermon, Aditya Grover, and Volodymyr Kuleshov, three professors whose academic research had focused on diffusion models and adjacent generative-AI techniques. Ermon's group at Stanford had been one of the early academic centers for diffusion-model research, with research contributions across diffusion architectures, flash-attention optimization, decision transformers, and direct preference optimization. Grover and Kuleshov had been Ermon's doctoral students and later established their own research groups at UCLA and Cornell respectively.

The founding thesis was that diffusion architectures, which had become the dominant approach for image and video generation through models such as Stable Diffusion and the broader latent-diffusion family, could be adapted to text and code generation with substantial throughput advantages over the autoregressive transformer architectures used by every major frontier large-language model. Diffusion-based large-language models (dLLMs), in the Inception framing, generate entire blocks of text in parallel through iterative refinement rather than predicting one token at a time, producing dramatic speedups for many real-world workloads.

In February 2025, Inception released Mercury, described as the world's first commercial-scale diffusion large-language model. The launch included a general-purpose Mercury model and a Mercury Coder variant tuned for code generation, both with 128,000-token context windows. In November 2025, Inception announced a $50 million seed round led by Menlo Ventures, with backing from a senior syndicate of venture-capital firms and angel investors including Andrew Ng and Andrej Karpathy. In February 2026, Inception launched Mercury 2, positioned as substantially faster than competing speed-optimized large-language models from OpenAI, Anthropic, and Google.

The company is distinct from G42's Inception division (an Abu Dhabi-based unit profiled separately at /inception/). The two share a name but are unrelated entities.

Mission and strategy

Inception Labs's stated mission is to make large-language models "10x faster and more efficient" through diffusion-based generation. Ermon has framed the technical premise: "These diffusion-based LLMs are much faster and much more efficient" because they process operations in parallel rather than sequentially.

The strategy combines three threads. First, foundational research on diffusion-based generation for text and code, an architectural alternative to the dominant autoregressive transformer family. Second, a developer-and-enterprise commercial focus, with Mercury distributed via API and integrated into developer tooling rather than positioned as a consumer chatbot. Third, partnerships with developer-platform companies such as ProxyAI, Buildglare, Kilo Code, OpenRouter, and Poe to embed Mercury throughput advantages directly into existing developer workflows.

The competitive premise is that throughput-and-cost differentiation can produce a durable commercial position even where Mercury does not lead generalist capability benchmarks. Mercury's reported pricing of $0.25 per million input tokens and $1.00 per million output tokens places it at substantially below the API pricing of frontier-tier autoregressive models from OpenAI and Anthropic. The combined pricing-and-speed advantage is positioned as the principal commercial wedge.

Models and products

• Mercury (February 2025). First commercial-scale diffusion large-language model. Available in general-purpose Mercury and Mercury Coder variants. Both feature 128,000-token context windows.
• Mercury 2 (February 2026). Refreshed flagship with quality improvements across coding, instruction following, mathematical reasoning, and knowledge recall, measured as the average of GPQA Diamond, LCB, IFBench, AIME 2025, and SciCode benchmarks. Reported throughput exceeding 1,000 tokens per second.
• API platform. Direct API access at $0.25 per million input tokens and $1.00 per million output tokens. Inception is among the lower-priced commercial-API providers as of April 2026.
• Partner integrations. Mercury is integrated into developer tools including ProxyAI, Buildglare, and Kilo Code, and is available via aggregator platforms OpenRouter and Poe.

The company has not released open-weights versions of the Mercury family as of April 2026.

Benchmarks and standing

Mercury 2 is positioned by Inception as the fastest commercially available large-language model rather than the highest-quality on capability benchmarks. The reported throughput of more than 1,000 tokens per second compares favorably with autoregressive baselines that typically operate at 30 to 200 tokens per second for similar response lengths.

On capability, Mercury 2 quality is measured as the average of GPQA Diamond, LCB, IFBench, AIME 2025, and SciCode. Specific scores have not been comprehensively published as of April 2026, and Mercury does not lead any major generalist capability leaderboard. The competitive position is throughput-and-cost rather than capability-leadership.

The combination of diffusion-architecture novelty, the Stanford-UCLA-Cornell academic-founder team, and the senior venture-capital syndicate produces a research-credibility profile that distinguishes Inception within the broader cohort of mid-size 2024-vintage Insurgent labs.

Leadership

As of April 2026, Inception Labs's named leadership includes:

• Stefano Ermon, co-founder and chief executive. Stanford professor of computer science whose research group has been one of the early academic centers for diffusion-model research. Public face for the company on technical claims and product framing.
• Aditya Grover, co-founder. UCLA professor. Former Ermon doctoral student, with research contributions across diffusion architectures and decision transformers.
• Volodymyr Kuleshov, co-founder. Cornell professor. Former Ermon doctoral student, with research contributions on diffusion-based generation and adjacent areas.

The senior research and engineering team has not been disclosed by name. The founder-professor structure is unusual within the cohort of well-funded 2024-vintage AI labs, the closest peer being the AAI lab founded by Mobileye chief executive Amnon Shashua and his Hebrew University-affiliated collaborators.

Funding and backers

Inception Labs's funding history through April 2026 consists of a single closed round: the $50 million seed announced in November 2025. The round was led by Menlo Ventures, with Mayfield, Innovation Endeavors, Microsoft M12, Snowflake Ventures, Databricks Ventures, and Nvidia NVentures participating. Angel investors included Andrew Ng and Andrej Karpathy.

The investor list combines two strategic patterns. The corporate-venture participants (Microsoft M12, Snowflake, Databricks, Nvidia) signal interest from large enterprise-AI platforms that could become distribution channels for Mercury. The angel-investor participation from Ng and Karpathy provides academic-research credibility within the broader AI ecosystem, given both are widely respected senior figures in the field.

The company has not disclosed cumulative compute commitments, cloud-partner relationships beyond the corporate-venture investors, or follow-on financing plans. The reported valuation of approximately $500 million is consistent with mid-size 2024-vintage Insurgent labs but trails the billion-dollar-plus seed valuations of Thinking Machines Lab, Humans&, and Flapping Airplanes.

Industry position

Inception Labs occupies a distinctive position within the 2024-vintage Insurgent cohort. The combination of an academic-professor founder team, the diffusion-architecture technical thesis, the developer-and-enterprise commercial focus, and the throughput-and-cost competitive premise produces a profile that does not directly mirror any other lab.

The closest peer comparators are research-first Insurgents that pursue novel large-language-model architectures or training methods. Liquid AI shares the academic-founder profile and the thesis that alternative architectures can outperform autoregressive transformers on certain workloads. RWKV and the broader open-architecture research community pursue related architectural alternatives.

The strategic risks are substantial. Diffusion-based large-language models have not yet been validated at the largest scales achieved by autoregressive frontier models. The throughput advantage may diminish as autoregressive labs deploy speculative decoding, parallel decoding, and similar techniques. The capability gap on generalist benchmarks limits direct competition with frontier models on quality-leading workloads.

The strategic strengths are equally distinctive. The Ermon-Grover-Kuleshov academic-research group is among the deepest diffusion-model research centers in the world, providing an architectural-research advantage that is difficult to replicate. The throughput-and-cost combination has demonstrated commercial traction through partner integrations. The corporate-venture investor list provides paths to enterprise distribution.

Competitive landscape

Inception Labs competes with several Frontier and Insurgent labs and adjacent product-layer companies:

• OpenAI and Anthropic. The dominant Frontier labs whose autoregressive models Mercury 2 is positioned against on throughput. Inception does not compete on generalist capability leadership.
• Mistral AI and Cohere. Mid-tier Frontier labs with developer-and-enterprise commercial focus. Mistral pursues open-weights distribution; Cohere pursues retrieval-augmented generation. Both compete with Inception for enterprise developer attention.
• DeepSeek. Chinese open-weights frontier capability at low per-token cost. The cost gap creates substitution risk within Inception's price-sensitive customer base.
• Liquid AI. Closest peer Insurgent on architecture-research grounds. Both pursue alternatives to autoregressive transformers, both founded by academic researchers.
• Cerebras and Groq. Hardware-accelerator companies that achieve high-throughput large-language-model inference through specialized silicon rather than architectural alternatives. Compete with Mercury on the throughput-and-cost axis through a different mechanism.

Outlook

Several open questions affect Inception Labs's trajectory in 2026 and 2027:

• The release timing and capability profile of any Mercury 3 generation, particularly whether the diffusion architecture can close the generalist-capability gap with frontier autoregressive models.
• Continued partner-integration adoption, particularly within developer-tooling and enterprise platforms.
• Pricing dynamics as autoregressive frontier models incorporate speculative decoding, parallel decoding, and similar throughput optimizations.
• Whether Inception accepts follow-on capital at a higher valuation, and on what timeline.
• Open-weights release plans, if any.
• Senior-talent recruitment from the Stanford-UCLA-Cornell research orbit and the broader academic ML community.

Sources

• TechCrunch: Inception raises $50 million to build diffusion models for code and text. Primary source on the $50 million seed round.
• BusinessWire: Inception Raises $50M to Power Diffusion LLMs. Investor list and capability claims.
• Inception Labs: The Next Step for dLLMs: Scaling up Mercury. Official Mercury 2 launch announcement and benchmark framing.
• Inception Labs: Introducing Mercury. Mercury 1 launch announcement.
• BusinessWire: Inception Launches Mercury 2, the Fastest Reasoning LLM. Mercury 2 launch coverage.
• Mayfield: Introducing Inception Labs. Investor profile of the company.
• Mercury: Ultra-Fast Language Models Based on Diffusion (arXiv preprint). Technical paper on the Mercury architecture.