Inception · 2026-09-08 · major
Mercury 2.5 — Inception's diffusion model hits 1,107 tokens per second
Mercury 2.5 is Inception's new diffusion language model, which the company calls the largest ever trained. It runs at 1,107 tokens per second on NVIDIA GPUs, handles 260K tokens of context, and costs $0.20 per million input tokens.
The largest diffusion language model trained so far, streaming over 1,100 tokens a second.
Key specs
| Throughput | 1,107 tokens/sec |
|---|
Quick facts
| Maker | Inception |
|---|---|
| Model type | Diffusion LLM (dLLM) |
| Context window | 260,000 tokens |
| Model ID | inception/mercury-2.5 |
| Availability | Inception API, Baseten, OpenRouter |
| What's new | 40% more intelligence than Mercury 2 |
| Released | September 8, 2026 |
Pricing
| Input | $0.20 / 1M tokens |
|---|---|
| Output | $0.75 / 1M tokens |
| Input (launch) · 80% launch discount | $0.04 / 1M tokens |
| Output (launch) · 80% launch discount | $0.15 / 1M tokens |
What is it?
Mercury 2.5 raises Inception's diffusion LLM line to what the company calls the largest ever trained, with a 40% gain in intelligence over Mercury 2. Diffusion language models write differently from the usual autoregressive kind: instead of producing one token after another, they start from a rough draft of the whole answer and refine tokens in parallel. Inception was founded by the Stanford, UCLA and Cornell researchers behind the first diffusion LLM.
How does it work?
Parallel refinement is what buys the speed. Because a dLLM improves many tokens at once rather than waiting for each previous token to land, Mercury 2.5 reaches 1,107 tokens per second on widely available NVIDIA GPUs. The release also adds tunable reasoning, parallel tool calls and schema-aligned JSON, and stretches the context window to 260K tokens.
Why does it matter?
Speed at a low price is the pitch. Inception rates Mercury 2.5's quality as comparable to cost-optimized frontier models while charging $0.20 per million input tokens and $0.75 per million output tokens, with an 80% discount at launch. For latency-sensitive work such as voice agents, inline code completion or high-volume classification, a model that streams over a thousand tokens a second removes a real bottleneck.
Who is it for?
developers building latency-sensitive apps
Frequently asked questions
- How does Mercury 2.5 compare to Mercury 2?
- Inception measures Mercury 2.5 at a 40% increase in intelligence over Mercury 2, and describes it as the largest diffusion language model ever trained. The company positions Mercury 2.5 as both its most capable diffusion LLM and the fastest reasoning model it has put into production.
- Which models does Inception compare Mercury 2.5 against?
- Inception places Mercury 2.5's quality alongside cost-optimized frontier models, naming GPT-5.6 Luna (Low), Gemini 3.5 Flash-Lite and Claude Haiku 4.5 as the comparison set. The claim is about the cost-optimized tier rather than the flagship tier, so Mercury 2.5 is pitched on the speed-and-price axis rather than on raw capability.
- Where can I run Mercury 2.5?
- Mercury 2.5 is available through the Inception API, Baseten and OpenRouter, where the model ID is inception/mercury-2.5. Inception also offers dedicated capacity deployments for enterprise customers. The release is served through APIs rather than as downloadable weights.
- What can Mercury 2.5 do besides plain text generation?
- Mercury 2.5 adds tunable reasoning, parallel tool calls and schema-aligned JSON output. Those features matter for agent workloads, where a model needs to call several tools in one turn and return data that fits a fixed schema. The 260K-token context window covers long documents and long agent traces.
Try it
Model ID: inception/mercury-2.5 on OpenRouter