Inception has released Mercury 3, a diffusion language model capable of generating text at 1,068 tokens per second on NVIDIA GPUs. The system accepts a 1 million token context window and claims to outperform GPT-6 Luna and Claude Haiku 5.5 on eight of twelve tests according to Inception’s own benchmarking table.
Core specs and availability
The model parameters remain undisclosed, though Inception states the size is more than 40 times larger than Mercury 1. The context window sits at 1 million tokens. Access is restricted to an API; there are no open weights or quantised builds available. The inference runs on Inception-served NVIDIA GPUs, with the specific hardware model unlisted.
Performance against rivals
On the SWE-Bench Atlas test, Mercury 3 scores 51.08, placing it approximately 18 points ahead of both competitors. However, the model trails Claude Haiku 5.5 on the DeepSWE v1 benchmark, scoring 46.31 against Haiku’s 51.62.
What is Mercury 3?
This release marks the fourth iteration of the Mercury series, following Mercury 1, Mercury 2, and Mercury 2.5. The initial Mercury 1 launched in 2025. Mercury 2.5 introduced tunable reasoning capabilities and expanded the context window to 260,000 tokens.
1/4 Today, we’re releasing Mercury 3. Mercury 3 is by far our most powerful model and the most intelligent diffusion LLM ever released. It outperforms GPT-6 Luna and Haiku 5.5 on quality while running at 1000+ tok/sec. pic.twitter.com/TYevDl43Lm
— Stefano Ermon (@StefanoErmon) October 19, 2026
How diffusion differs from standard generation
Autoregressive models produce text one token after another, forcing each new word to wait for the previous one to finish. A diffusion LLM begins with a rough draft of the entire response. It then refines every position simultaneously over several passes.
Reducing sequential steps allows for faster output on identical hardware. The central question for researchers was whether this parallel approach remains effective as models grow in size.
Does diffusion scale?
Inception’s data suggests yes. Mercury 3 contains more than 40 times the parameters of Mercury 1. Average quality across standard benchmarks increased by roughly three times between those two versions. Speed remained above 1,000 tokens per second.
The Inception team illustrates this quality versus size curve across five Mercury generations.
Benchmark results
The scores below derive from Inception’s table, with all models running at medium settings.
Mercury 3 shows its largest improvements in instruction following and agentic coding. The IFBench score reaches 76.84, which is 15 points higher than GPT-6 Luna. SWE-Bench Atlas hits 51.08, compared to 33.33 for Luna and 32.26 for Haiku 5.5. GPQA Diamond lands at 89.23.
Specific weaknesses exist. Haiku 5.5 wins DeepSWE v1, scoring 51.62 to Mercury 3’s 46.31. It also beats Mercury 3 on MRCR v2, with 78.40 against 75.50. GPT-6 Luna edges ahead on AA-Omniscience-Public (46.94 vs 46.44) and AA-LCR v1.1 (78.30 vs 77.67). Long-context retrieval and extended coding tasks remain difficult areas.
Comparison table
The following table compares Mercury 3 against GPT-6 Luna and Claude Haiku 5.5.
| Feature | Mercury 3 | GPT-6 Luna | Claude Haiku 5.5 |
|---|---|---|---|
| Developer | Inception | OpenAI | Anthropic |
| Generation method | Diffusion (parallel refinement) | Not disclosed | Not disclosed |
| Parameters | Not disclosed | Not disclosed | Not disclosed |
| Context window | 1M | 1.05M | 1M |
| Max output | Not disclosed | 128K | 128K |
| Input modalities | Not disclosed | Text, image | Text, image |
| Open weights | No (API) | No (API) | No (API) |
| Output speed (Inception’s measure) | 1,068 tok/s | ~130 tok/s | ~240 tok/s |
| Price, $/1M in / out | $0.20 / $0.70 list ($0.10 / $0.35 launch) | $0.10 / $0.50 (short context) | $0.10 / $0.50 (prompts up to 100K) |
| GPQA Diamond | 89.23 | 86.36 | 84.51 |
| IFBench | 76.84 | 61.82 | 54.65 |
| SWE-Bench Atlas | 51.08 | 33.33 | 32.26 |
| DeepSWE v1 | 46.31 | 37.76 | 51.62 |
| MRCR v2 | 75.50 | 62.72 | 78.40 |
Speed and benchmark rows: Inception’s table, medium settings. Specs and prices: each vendor’s own docs.
Early user reports
Inception cites four early-access partners. These are customer statements, not independent evaluations.
- TinyFish ran 50 live Mind2Web web-navigation tasks. Mercury 3 completed 50% of them. Inception claims this beats GPT-6.1-Sol, though no Sol score is provided.
- Octen AI reports a median generation time of 0.9 seconds and 2.2 seconds end-to-end for smart-glasses search.
- RevisionDojo notes that hints and marked answers return in about a second.
- Lorikeet states Mercury 3 beat other models on latency and evaluations for support agents.
Developer access and pricing
Mercury 3 is live on the Inception API Platform, OpenRouter, Baseten, and Vercel. The API is OpenAI compatible. Documentation uses the model string mercury-3 and a reasoning_effort parameter.
List pricing is $0.20 per 1 million input tokens and $0.70 per 1 million output tokens. At launch, the price is 50% off at $0.10 and $0.35. Cached input receives a 90% discount. New accounts get 100 million free tokens. Enterprise plans add dedicated capacity, autoscaling, and configurable data retention.
Inception also ships two siblings:
- Mercury Voice targets time to first answer token in voice agents.
- Mercury Decide is a classification model.
What it means
The practical result for developers is a tool that generates code and text significantly faster than standard models. The speed allows for real-time interaction in voice bots and search tools where latency was previously a barrier. While the model struggles with very long coding tasks compared to Haiku 5.5, the throughput means agents can execute more complex loops without the user waiting for each line to appear.
