vs
StepFun vs Inception
StepFun makes efficient multimodal models with open Apache 2.0 weights. Inception makes closed text models built for raw speed.
By The Subconscious Team · Updated
StepFun vs Inception: key differences
StepFun's Step 3.7 Flash is a 198B mixture-of-experts vision-language model with 11B active, 256K context and tool use, priced at $0.20 in and $1.15 out on StepFun's API, at about 128 tokens per second. Inception sells closed Mercury diffusion models that refine a block of tokens in parallel instead of one at a time. Mercury 2.5 is a reasoning model with tool use and 260K context at $0.20 in and $0.75 out per million tokens; Inception reports over 1,100 tokens per second, and Artificial Analysis measures about 640.
Pick StepFun when the workload needs images or video, or when you want weights you can self-host. Pick Inception when output speed drives the product, as in autocomplete, voice agents or real-time chat, and text input is enough.
What StepFun and Inception do
StepFun
StepFun is a Shanghai AI lab known for efficient multimodal models, with a mix of proprietary API models and open-weight releases. In September 2026 it launched Step 5 Preview, its new flagship for agentic work: a 600B-total, 27B-active mixture-of-experts model with 1M context and text and image input, which Artificial Analysis scores level with Kimi K3. Its cheaper workhorse, Step 3.7 Flash, came out in May 2026 as a 198B mixture-of-experts vision-language model with only 11B active parameters. It has 256K context, selectable reasoning levels, tool use and structured outputs, and it ships under Apache 2.0. StepFun's own API prices it at $0.20 in and $1.15 out per million tokens, and OpenRouter carries it too.
Example models: Step 5 Preview, Step 3.7 Flash
Full StepFun profile
Inception
Inception builds diffusion large language models (dLLMs). Every other LLM on this list generates text one token at a time; Inception's Mercury models start from a rough draft of a whole block and refine it over a few passes, producing many tokens in parallel. That ties speed less to output length and lets the models run fast on ordinary NVIDIA GPUs rather than custom chips. The company was founded by professors from Stanford, UCLA and Cornell behind core work on diffusion, Flash Attention and DPO, and is led by Stanford's Stefano Ermon. It raised a $50M seed in November 2025 led by Menlo Ventures, with Mayfield, M12, NVentures, Snowflake and Databricks, plus Andrew Ng and Andrej Karpathy as angels.
Example models: Mercury 2.5, Mercury 2
Full Inception profileShould you choose StepFun or Inception?
StepFun vs Inception at a glance
| Attribute | StepFun |  Inception |
|---|---|---|
| Model access | Open (Apache 2.0) and API models | Closed weights (Mercury dLLMs) |
| Flagship models | Step 5 Preview, Step 3.7 Flash | Mercury 2.5, Mercury 2 |
| Speed | ~128 tok/s on Step 3.7 Flash | 1,100+ tok/s (vendor); ~640 measured |
| Price | $0.20 in, $1.15 out (Step 3.7 Flash) | $0.20 in, $0.75 out (Mercury 2.5) |
| Customization | Open weights to fine-tune | Fine-tuning on request |
| Deployment | First-party API, OpenRouter | Serverless API, Bedrock, Azure, private |
| Long context | Up to 1M (Step 5 Preview); 256K on Step 3.7 Flash | 260K |
Frequently asked questions
What is the difference between StepFun and Inception?
StepFun makes efficient multimodal models with open Apache 2.0 weights. Inception makes closed text models built for raw speed.
When should I choose StepFun over Inception?
Image and video understanding; Apache 2.0 weights to self-host; Low-cost multimodal agents.
When should I choose Inception over StepFun?
Very fast output on standard GPUs; Cheap reasoning with tool use and 260K context; Buying through AWS Bedrock or Azure.
Is StepFun or Inception cheaper?
StepFun: $0.20 in, $1.15 out (Step 3.7 Flash). Inception: $0.20 in, $0.75 out (Mercury 2.5). The cheaper choice depends on the model and workload.
Which has more context, StepFun or Inception?
StepFun: Up to 1M (Step 5 Preview); 256K on Step 3.7 Flash. Inception: 260K.
Related comparisons
Running long-horizon agents?
If your agents run past 200K tokens, compare both against Subconscious. Our inference stack treats a long-horizon trace as the primary workload, so speed, cost, and accuracy hold up deep into the trace.