DeepInfra raises $107M Series B to scale the inference cloud — read the announcement

Step 3.7 Flash is Live on DeepInfra: An Agentic, Multimodal Model Built for Production

Published on 2026.06.12 by DeepInfra
Step 3.7 Flash is Live on DeepInfra: An Agentic, Multimodal Model Built for Production

Step 3.7 Flash from StepFun is now live on DeepInfra. It's a 198B-parameter sparse Mixture-of-Experts vision-language model that activates only about 11B parameters per token, supports a 256K context window, and exposes three selectable reasoning levels—so you can dial the trade-off between speed, cost, and depth on a per-request basis. We've deployed it at the same competitive pricing you'll find elsewhere, and it's available through our standard OpenAI-compatible API with no special setup.

Step 3.7 Flash wasn't built to win isolated benchmarks. It was built for developers scaling agentic workflows that combine perception, search, and reasoning—parsing a massive financial report in a single pass, running multi-step search loops with cross-source verification, or operating concurrent coding agents in a high-throughput pipeline. That focus shows up everywhere in how the model behaves.

Built for Agentic Workflows

For autonomous agents, execution reliability matters more than raw model quality. An agent that drifts from instructions, violates a system constraint, or falls for an adversarial trap mid-trajectory is worse than useless. Step 3.7 Flash is tuned for exactly this kind of long-horizon, multi-turn orchestration.

  • ClawEval-1.1: 67.1 — leading the benchmark and well ahead of the next-closest competitor, reflecting strong resistance to adversarial traps and strict adherence to system policy across multi-turn runs.
  • Toolathlon: 49.5 and HLE with Tool: 47.2 — dependable tool orchestration, calling external APIs and executing long workflows without losing the thread.
  • SWE-Bench Pro: 56.3 — a second-place finish on live engineering tasks: tracing multi-file repositories, isolating bugs from raw issue reports, and generating patches that pass automated tests.

Multimodal Perception

Step 3.7 Flash pairs its 196B-parameter language backbone with a 1.8B-parameter vision encoder for native image understanding—not bolted-on captioning, but visual grounding that feeds directly into reasoning and retrieval.

  • SimpleVQA (Search): 79.2 — first place, combining visual grounding with retrieval-augmented verification.
  • V* (Python): 95.3 — frontier parity on fine-grained visual reasoning.

In practice that means the model reads dense visual interfaces—UI wireframes, application GUIs, data charts—and maps them into structured code. When a visual asset is incomplete, it can recognize what's missing, run a lookup to fill the gap, and verify its conclusion before answering.

Three Reasoning Levels

Step 3.7 Flash exposes low, medium, and high reasoning levels through a single reasoning_effort parameter. Use low for latency-sensitive, high-volume calls; reach for high when a task needs deeper deliberation. It's the same model and the same endpoint—you just choose how much thinking to spend per request.

Pricing

Token typePrice per 1M tokens
Input (cache miss)$0.20
Input (cache hit)$0.04
Output$1.15

The aggressive cache-hit rate rewards the prefix-heavy prompts that agentic and multi-turn workloads naturally produce.

Getting Started

Step 3.7 Flash is available today through DeepInfra's OpenAI-compatible API. If you've used DeepInfra before, nothing changes—same API, same setup. Point your client at our endpoint and use stepfun-ai/Step-3.7-Flash as the model name:

from openai import OpenAI

client = OpenAI(
    api_key="$DEEPINFRA_TOKEN",
    base_url="https://api.deepinfra.com/v1/openai",
)

# Text + image input, with a chosen reasoning level
completion = client.chat.completions.create(
    model="stepfun-ai/Step-3.7-Flash",
    reasoning_effort="medium",
    messages=[
        {
            "role": "user",
            "content": [
                {"type": "text", "text": "What's in this chart, and what does it imply?"},
                {"type": "image_url", "image_url": {"url": "https://example.com/chart.png"}},
            ],
        }
    ],
)

print(completion.choices[0].message.content)
copy

See Step 3.7 Flash for live pricing and usage, browse the full model catalog, or read the docs to start building.

Related articles
DeepSeek V4 Pro 0813 on DeepInfra: Pricing, Benchmarks & APIDeepSeek V4 Pro 0813 on DeepInfra: Pricing, Benchmarks & API<p>Updated November 2026 for the DeepSeek-V4-Pro-0813 release. DeepSeek&#8217;s GA release of V4 Pro moves its DeepSWE score from 12.8 to 62.7 and Terminal Bench 2.1 from 72.1 to 87.9, according to DeepSeek&#8217;s model card (vendor-reported). DeepSeek-V4-Pro-0813 shipped on August 13, 2026 and supersedes the April preview. It keeps the same 1.6-trillion-parameter Mixture of Experts design [&hellip;]</p>
Model Distillation Making AI Models EfficientModel Distillation Making AI Models EfficientAI Model Distillation Definition & Methodology Model distillation is the art of teaching a smaller, simpler model to perform as well as a larger one. It's like training an apprentice to take over a master's work—streamlining operations with comparable performance . If you're struggling with depl...
OpenClaw Cost Optimization: Cut AI API Costs by 90%OpenClaw Cost Optimization: Cut AI API Costs by 90%<p>A single ask in an OpenClaw session can cost more than a full evening of casual ChatGPT use. Ask your agent something simple, like which calendar event clashes with your flight, and the request that hits the API carries far more than your 12-token question. It also carries your SOUL.md, the tool schemas registered on [&hellip;]</p>