Inception Labs' Mercury 3 Hits 981 Tokens per Second on Reasoning Tasks

Inception's Mercury 3 Preview became the first diffusion LLM scored on ARC-AGI, averaging 981 tokens/sec and peaking at 1,743.

·
·
·
Inception Labs' Mercury 3 Hits 981 Tokens per Second on Reasoning Tasks
Read5 min
TypeNews
SubtopicLong Context
  • ARC Prize verified Inception's Mercury 3 Preview, the first diffusion LLM it has ever evaluated.
  • Scores: 88.5% on ARC-AGI-1 at $0.03/task, 61.8% on ARC-AGI-2 at $0.06/task.
  • ARC-AGI-3 was a near-zero result (0.1-0.2%) at $299-$333 per task.
  • Averaged 981 output tokens/sec on ARC-AGI-2, peaked at 1,743 tokens/sec in a single call.
  • Mercury 3 offers a 1M-token context window at $0.10/$0.35 per million input/output tokens.
  • Available via OpenRouter, Baseten, Vercel, and Inception's own API.

ARC Prize has published verified results for Mercury 3 Preview, the first diffusion language model evaluated on its ARC-AGI suite. The model scored 88.5% on ARC-AGI-1 and 61.8% on ARC-AGI-2 while averaging 981 output tokens per second on the public ARC-AGI-2 dataset under its fastest setting.

Mercury 3 Preview comes from Inception Labs, a Redwood City company developing commercial diffusion LLMs. ARC, short for Abstraction and Reasoning Corpus, tests whether models can infer rules from sparse examples and apply them to unfamiliar tasks. The verification provides an external measure of how diffusion decoding performs on reasoning workloads.

The scores diverge on interactive tasks

Verified Mercury 3 Preview results
Benchmark Score Cost per task
ARC-AGI-1 88.5% $0.03
ARC-AGI-2 61.8% $0.06
ARC-AGI-3, provider adapter 0.2% $299
ARC-AGI-3, standard 0.1% $333

ARC-AGI-1 and ARC-AGI-2 use static visual-grid puzzles. A model receives a few input-output examples, infers the transformation, and produces the correct output grid for a new input. ARC-AGI-3 changes the format to interactive game environments that require exploration, state tracking, and repeated actions. Each benchmark generation also changes task difficulty, so cross-version percentage comparisons have limited meaning.

Mercury 3 Preview performed well on the static puzzles at a few cents per task. Its ARC-AGI-3 score remained at or below 0.2% in both published configurations. Repeated calls across the interactive turns pushed the cost to $299 and $333 per task, exposing a poor fit for long-horizon agent work under these settings.

Nearly 1,000 tokens per second

With reasoning effort set to “instant,” Mercury 3 Preview averaged 981 output tokens per second across the public ARC-AGI-2 dataset. That configuration used no reasoning tokens. The fastest cited request reached 1,743 tokens per second, generating 3,001 tokens in 1.72 seconds.

Inception reports 1,068 tokens per second on NVIDIA Blackwell GPUs in broader testing and claims a fourfold to eightfold advantage over comparable autoregressive systems. The ARC result supports the company’s throughput figures, although cross-model comparisons also depend on hardware, batch size, tokenizer, output length, and serving configuration.

Output throughput captures only part of application latency. Production systems must also account for prompt processing, time to first token, network delays, concurrency, tool execution, and retries. Those factors can dominate short responses even when sustained generation is fast.

Parallel refinement cuts serial work

Autoregressive language models generate tokens sequentially from left to right, with each new token depending on the preceding sequence. A diffusion language model begins with a block of masked or noisy tokens and refines multiple positions during each pass. Resolving several positions at once reduces the number of serial decoding steps.

The model still performs repeated refinement passes, and the amount of computation depends on its architecture and decoding settings. Parallel refinement becomes especially useful for long outputs, where sequential generation creates a larger latency penalty.

ARC-style workloads can involve lengthy candidate solutions or reasoning traces. Faster decoding can reduce the wall-clock cost of repeated sampling, self-consistency checks, and agent loops. Mercury’s ARC-AGI-3 result also shows that faster generation cannot compensate for weak planning or ineffective exploration.

Preview results, product details

ARC evaluated an endpoint named Mercury 3 Preview. Inception’s broader Mercury 3 release materials list the following product specifications, which developers should confirm against the preview endpoint they plan to use:

  • Context: Up to 1 million tokens, increased from 260,000 in Mercury 2.5.
  • Pricing: Launch rates of $0.10 per million input tokens and $0.35 per million output tokens.
  • Output: Up to 131,072 tokens, with configurable reasoning effort, tool calling, and structured outputs.
  • Scale: More than 40 times as many total parameters as Mercury 1, with similar reported generation speed.
  • Access: Inception’s API platform, OpenRouter, Baseten, and Vercel, with an OpenAI API-compatible interface.
  • Licensing: Closed and proprietary, with no published weights for self-hosting or fine-tuning.

Workloads that match the profile

Inception says early-access partners are using Mercury 3 for browser agents, interactive education, real-time search, and customer support. The ARC results support a narrower conclusion about fast generation and static reasoning, so each production workload still requires its own quality and latency tests.

  • Code and document generation: High decoding throughput can shorten long edits, transformations, and structured responses.
  • Voice and support agents: Faster generation can reduce response time, although short exchanges often depend more on time to first token and the surrounding speech pipeline.
  • Browser and computer agents: Structured action plans may benefit from fast output, while the ARC-AGI-3 result raises concerns about extended exploration and recovery from mistakes.
  • Long-context retrieval: The advertised 1 million-token window can accommodate large corpora, but ARC does not measure retrieval accuracy, prompt-processing time, or performance across the full window.
  • Autonomous exploration: Near-zero ARC-AGI-3 scores indicate that this preview and configuration are currently a weak choice for unfamiliar, multi-step environments.

What production tests still need to cover

The verified run provides a useful throughput and reasoning data point, while deployment decisions require measurements from the intended stack. A practical evaluation should cover:

  • Median and tail latency, including time to first token.
  • Quality and cost at each reasoning-effort setting.
  • Tool-call accuracy and structured-output compliance.
  • Prompt-processing time for large context windows.
  • Performance under concurrent load, retries, and rate limits.
  • Task success across long agent loops, not only per-turn output quality.

ARC’s verification shows that a commercial diffusion LLM can solve difficult static reasoning tasks while sustaining close to 1,000 output tokens per second in this run. Inception’s broader claims about diffusion scaling to frontier-level general reasoning remain open, particularly for interactive agents. For developers whose workloads are dominated by long generation, Mercury 3 Preview offers a distinct latency profile worth testing against end-to-end quality, reliability, and cost requirements.

Trending
  • No trending articles

Comments

avatar

Next Reads