Mercury 3 Hits 1,000 Tokens a Second, Testing Diffusion at Scale
Inception's Mercury 3 claims 1,000+ tokens per second and benchmark wins over GPT-6 Luna and Claude Haiku 5.5, but diffusion's serving edge under large production batches remains unproven.
Inception Labs has released Mercury 3, a diffusion-based language model it says outperforms GPT-6 Luna and Claude Haiku 5.5 on standard benchmarks while generating over 1,000 tokens per second, 4-8x faster than those models.
The company reports 1,068 tokens per second on NVIDIA GPUs and a 1M-token context window, up from 260K for Mercury 2.5, at launch pricing of $0.10 per million input and $0.35 per million output tokens, half the list rate, with cached tokens 90% off input price. Inception attributes the speed to parallel iterative denoising rather than one-token-at-a-time generation, arguing the advantage grows as reasoning models produce longer answers.
The bits that matter
- Inception reports Mercury 3 completed 50% of 50 live Mind2Web navigation tasks inside its production web-agent harness, more than GPT-6.1-Sol, using the same token volume and agent steps at a fraction of the per-token cost.
- Early-access partners describe latency-sensitive deployments: RevisionDojo says hints, marked answers and definitions return in about a second, while Octen AI reports a median generation time of 0.9 seconds and 2.2 seconds end to end.
- Mercury 3 is available on the Inception API Platform, OpenRouter, Baseten and Vercel, and is OpenAI API compatible; enterprise deployments add dedicated capacity, autoscaling, compliance controls and configurable data retention.
- Inception also launched Mercury Voice for low-latency voice applications and Mercury Decide for high-volume classification, which it says outperforms Jev on JevBench and LegalBench.
Discussion analysis
Analysis details
7 comments analyzed. Analyzed .
The sample selects active discussion branches and includes available parent comments. It may omit parts of the full thread. The analysis included 7 of 8 usable stored comments because of the input limit. Selected themes do not measure community opinion or how common a view is.
Speed-first use cases remain narrow
The case for a 1,000-token-per-second model rests on workloads where latency compounds, such as browser agents, voice tutoring and real-time search. A skeptic argues it is hard to name tasks that genuinely require that speed, since the familiar trade-off is price against quality and reliability rather than raw throughput. That narrows the addressable case: Inception's own examples are latency-sensitive products, not general chat.
Launch discount shapes the adoption math
A prospective user plans to switch from Haiku and GPT-Luna for speed-sensitive work but questions whether Mercury 3 stays worthwhile once the 50% launch discount ends, when output tokens rise from $0.35 to $0.70 per million. The speed advantage then has to justify itself at list price rather than introductory pricing.
Batch size may blunt diffusion's serving edge
A reply about Google's stalled diffusion work notes that its preprint performed well at small batch sizes, while most production deployments serve many users at once with large batches. If diffusion's parallel denoising advantage narrows under high concurrency, single-stream throughput numbers may overstate the gain in shared serving.
Google's diffusion lead went quiet
Google's Gemini Diffusion demo drew attention for its speed but produced little visible follow-through, and a later Diffusion Gemma release made limited noise. The comparison raises the question of whether a fast diffusion demo converts into a supported product, which is the gap Inception now claims to fill.
Access gaps: Baseten listing and open weights
Two access questions sit alongside the launch. One asks why Mercury 3 is not visible on Baseten despite being named as a distribution partner. Another notes that no open-weights version has appeared, leaving self-hosting off the table for teams that want it.