Mercury 3 diffusion LLM reaches 1,068 tokens per second

What changed

Inception released Mercury 3 on October 19, 2026, the fourth generation in its diffusion-based language model series. The company reports the model achieves 1,068 tokens per second on NVIDIA GPUs with a 1M token context window, significantly expanded from Mercury 2.5’s 260K context. Inception states Mercury 3 contains more than 40 times the total parameters of Mercury 1, which launched in 2025.

On Inception’s self-reported benchmarks run at medium reasoning settings, Mercury 3 leads GPT-6 Luna and Claude Haiku 5.5 on 8 of 12 tests. The model scores 51.08 on SWE-Bench Atlas for software engineering tasks, approximately 18 points above both competitors, and 76.84 on IFBench for instruction following, about 15 points ahead of Luna. However, Haiku 5.5 still leads on DeepSWE v1 with 51.62 compared to Mercury 3’s 46.31, and trails Mercury 3 on long-context retrieval benchmarks like MRCR v2.

Launch pricing is set at $0.10 per million input tokens and $0.35 per million output tokens, offering a 50% discount from standard rates of $0.20 and $0.70. Cached input tokens receive a 90% discount. Mercury 3 launches today via the Inception API Platform, OpenRouter, Baseten, and Vercel with OpenAI-compatible API support.

Why it matters

This release tests whether diffusion models can scale to practical speeds while competing with mainstream autoregressive systems on production workloads. Teams building real-time applications like browser agents, voice bots, and interactive tutors face a trade-off between speed and reasoning quality. Mercury 3’s reported speed advantage potentially shifts that calculation in favor of faster responses without sacrificing capability on certain tasks.

For cost-sensitive teams, the speed gain addresses a key bottleneck: latency-sensitive workflows consume tokens faster than necessary when waiting on sequential model responses. An agent that pauses at each step compounds delays across multiple reasoning cycles. Agents receiving answers in milliseconds instead of seconds can complete tasks sooner and explore more problem-solving approaches before responding to users.

The benchmark results suggest that diffusion LLMs may finally achieve parity with speed-optimized frontier models on complex reasoning and code generation. Yet weak spots remain: Mercury 3 underperforms on long-context retrieval and extended coding tasks, indicating that diffusion LLMs still encounter structural limits when handling very large documents or extended prompts.

Four early-access partners have shared results. TinyFish completed 50 percent of 50 live web navigation tasks on Mind2Web, which Inception says exceeded GPT-6.1-Sol performance. Octen reported 0.9 seconds median generation time for smart-glasses search. RevisionDojo cited approximately one-second response times for tutoring hints and marked answers. Lorikeet noted Mercury 3 demonstrated better latency than competing models for support agents. These statements come from customers, not independent evaluators.

Speed gains also reduce infrastructure costs for high-volume use cases. A voice agent that saves milliseconds per request across thousands of concurrent users compounds into significant compute savings over time.

What to test

Before deploying Mercury 3 to production, teams should independently validate several key claims:

• Benchmark reproducibility. Run Mercury 3 alongside GPT-6 Luna and Claude Haiku 5.5 on your own reasoning and coding tasks using identical settings. Inception’s benchmarks are vendor-reported on tests the company selected. Your specific workload may show different performance profiles.

• Latency under realistic load. Test token generation speed not just at low concurrency but under peak loads matching your actual usage patterns and hardware configuration. GPU type, batch size, and concurrent requests all affect real-world throughput substantially.

• Context window quality at scale. The 1M token context is prominent in marketing but verify whether output quality degrades at extreme token positions. Test retrieval and reasoning tasks that exercise the full context window rather than relying on summarized examples.

• Long-context retrieval performance. Mercury 3 underperforms on MRCR v2 and related long-context retrieval benchmarks. If your application requires searching or reasoning over lengthy documents, test this scenario thoroughly before committing to production use.

• Total cost evolution. The 50% launch discount is temporary. Model the full cost including inference, cached tokens, and standard pricing once the promotional period expires. Compare total cost per completed query rather than focusing only on per-token rates.

• API reliability and support. Verify uptime guarantees, rate limits, and failover procedures before integrating Mercury 3 into production systems handling support or search workloads that users depend on.

• Data handling and compliance. Mercury 3 operates exclusively through Inception’s managed API with no open-source weights or local deployment option. Confirm data retention practices and compliance certifications align with your requirements, particularly if your industry requires healthcare or financial regulatory compliance.

Common questions

What is a diffusion LLM, and how is it different?

Autoregressive models generate one token sequentially after another, with each new token waiting for the previous one to complete. Diffusion LLMs start with a rough draft representing the complete answer and refine every position in parallel across multiple passes. Fewer sequential steps means faster output on the same hardware, though output quality depends on the number of refinement passes the model performs. Inception reports that Mercury 3’s approach maintains speed advantages while scaling to larger parameter counts.

Are diffusion models faster than standard LLMs?

Inception reports Mercury 3 runs 4 to 8 times faster than GPT-6 Luna and Claude Haiku 5.5, with throughput reaching 1,068 tokens per second on NVIDIA GPUs. Speed benefits vary by task type and hardware configuration. Long-context retrieval and extended reasoning still appear slower for diffusion compared to certain autoregressive competitors, suggesting speed advantages are task-dependent rather than universal.

Where can I use Mercury 3?

Mercury 3 is available exclusively as a managed API through Inception, OpenRouter, Baseten, and Vercel. No open-source version or quantized weights exist for local inference. The API is OpenAI compatible, allowing it to function as a drop-in replacement for existing code using OpenAI-compatible endpoints.

The conclusion

Mercury 3 offers meaningful speed improvements combined with competitive quality on benchmarks where Inception has evaluated it. The 1,068 tokens per second throughput and 1M context window make it a credible choice for teams building real-time chat, tutoring, search, and voice applications. Launch pricing at $0.10 and $0.35 per million tokens reduces the cost barrier for experimentation.

Significant risks warrant consideration. Inception conducted and reported all benchmarks internally rather than through independent evaluation. Long-context retrieval capabilities still lag competitors. The model operates solely through managed API with no self-hosted or local inference option. Inception has shipped four Mercury releases in under two years, demonstrating rapid iteration but raising questions about model stability and long-term support commitment.

For teams already invested in real-time AI and willing to accept dependence on a vendor’s API infrastructure, Mercury 3 merits trial deployment. For teams requiring long-context reasoning, local inference capabilities, or independently verified performance claims, open-weight alternatives offer reduced vendor dependency and greater control. Speed benefits matter most for agents and search systems where millisecond savings multiply across millions of production queries. Monitor whether Inception maintains output quality as the launch discount expires and standard pricing takes effect.

Related reading

• Drex Max: $0.06/M tokens, 67.59 on Decision Index
• MiniMax M3.1 Flash: $0.15/M tokens cuts agent costs 70%
• Claude Haiku 5.5 vs Gemini 4 Argon: Cost and Speed