Moondream 3 Preview vs Step 3.5 Flash: which should you choose in 2026?
Moondream 3 Preview (by Moondream) and Step 3.5 Flash (by StepFun) are compared below. Here is how they stack up on benchmarks, price, and capabilities, and which one to pick in 2026.
There are not enough shared, protocol-compatible benchmark results to declare a performance leader.
Step 3.5 Flash is about 5.7× cheaper on a blended 3:1 input/output basis ($0.15 vs $0.85 per 1M tokens).
Step 3.5 Flash has the larger context window (262,144 tokens vs 32,000 tokens).
Choose Moondream 3 Preview if…
- • you need image inputs
Choose Step 3.5 Flash if…
- • lower blended API cost matters for your workload
- • you work with longer documents, transcripts, or codebases
The benchmark count includes only results measured with a matching benchmark version and protocol. Arena scores are shown separately. Missing, preliminary, and incompatible data is not treated as a controlled win. Published point-score comparisons are labeled separately when protocol details are incomplete. For text-output models, the verdict also compares token pricing and context windows.
Performance benchmarks
Every value links to its source. A dash means that no reviewed result is available for that exact model and protocol.
| Benchmark | Moondream 3 Preview | Step 3.5 Flash |
|---|---|---|
A 657-task benchmark of multi-step work across simulated SaaS applications in six business domains. The Artificial Analysis protocol reports a guardrail-aware score and is distinct from both the public Zapier split and the unrelated dynamic AutoBench framework. Guardrail-aware score | — | — |
A benchmark of difficult, verifiable information-seeking questions designed to measure an agent's ability to locate hard-to-find facts through web browsing. Accuracy | — | — |
A verified computer-use benchmark in which multimodal agents operate desktop applications and are graded from the resulting environment state. Mean task reward | — | — |
A human-validated subset of real GitHub issues used to measure whether a coding agent can produce repository patches that resolve the associated tests. Resolved | — | — |
Version 2.1 of the benchmark for completing realistic tasks in terminal environments. Harness, resource limits, and attempt count are part of the protocol. Mean task success | — | — |
A dynamic LLM evaluation framework in which models generate questions, answer them, and participate in reciprocal peer assessment. AutoBench is distinct from Zapier's AutomationBench. Weighted peer-assessment score | — | — |
The highest-quality subset of Graduate-Level Google-Proof Q&A, designed to test expert-level scientific reasoning in biology, physics, and chemistry. Accuracy | — | — |
A 2,500-question expert-level benchmark spanning dozens of academic fields. Tool-assisted and no-tools results are separate protocols and must not be merged. Accuracy | — | — |
| Arena preference scores | ||
Human preference score | — | 1393 ±4 |
Benchmark sources
Reviewed evidence last updated 2026-08-09. Arena data last refreshed Oct 5, 2026.
Pricing, capabilities, and model facts
| Feature | Moondream 3 Preview | Step 3.5 Flash |
|---|---|---|
| Context & model facts | ||
| Developer | Moondream | StepFun |
| API provider | Moondream | StepFun |
| Input context | 32,000 tokens | 262,144 tokens |
| Maximum output | 4,096 tokens | 65,536 tokens |
| Released | — | — |
| Added to Writingmate | Oct 15, 2026 | Jan 29, 2026 |
| License | Not available | Not available |
| Knowledge cutoff | — | — |
| Capabilities | ||
| Inputs | Text, Image | Text |
| Outputs | Text | Text |
| Provider endpoint accepts tool parameters | No | Yes |
| Reasoning | Yes | Yes |
| Vision | Yes | No |
| Image Generation | No | No |
| Video Generation | No | No |
| API pricing | ||
| Input (per 1M tokens) | $0.30 | $0.10 |
| Output (per 1M tokens) | $2.50 | $0.30 |
| Blended 3:1 input/output | $0.85 | $0.15 |
| API performance | ||
| p95 latency | Not measured | Not measured |
| Output throughput | Not measured | Not measured |
| Writingmate shows API performance only when both models have enough observations from the same measurement window, prompt profile, and provider. Third-party latency values are not copied into this table. | ||
Data sources
- Writingmate model catalog (pricing, limits, and availability)
Catalog data last updated Oct 15, 2026.
Why Pay for Multiple Subscriptions?
Comparing Moondream 3 Preview from Moondream with Step 3.5 Flash from StepFun? Instead of managing separate API keys and subscriptions, get both with Writingmate.
| Plan | Price | Moondream 3 Preview | Step 3.5 Flash | AI Images | AI Video |
|---|---|---|---|---|---|
Writingmate Pro Most popular | $20/mo | Included | Included | Nano Banana Pro, FLUX.2, DALL-E & more | VEO 3.1, Kling 3.0 |
Writingmate Ultimate Power users | $60/mo | Included | Included | Nano Banana Pro, FLUX.2, DALL-E & more | VEO 3.1, Kling 3.0 |
Moondream 3 Preview vs Step 3.5 Flash FAQ
Which is better, Moondream 3 Preview or Step 3.5 Flash?
There are not enough shared, protocol-compatible benchmark results to declare a performance leader. Step 3.5 Flash is about 5.7× cheaper on a blended 3:1 input/output basis ($0.15 vs $0.85 per 1M tokens). Step 3.5 Flash has the larger context window (262,144 tokens vs 32,000 tokens).
Which model is cheaper to use through an API?
Step 3.5 Flash is about 5.7× cheaper on a blended 3:1 input/output basis ($0.15 vs $0.85 per 1M tokens).
Which model supports more context?
Step 3.5 Flash has the larger context window (262,144 tokens vs 32,000 tokens).
Can I switch between Moondream 3 Preview and Step 3.5 Flash?
Yes. Use the model selector on this page to open any current Writingmate model comparison. You can also run the same prompt with both models in Writingmate.
Related comparisons
More Moondream 3 Preview comparisons
DeepSeek V4.1 Flash 1017
DeepSeek
DeepSeek V4.1 Flash 1017 is the October 17, 2026 snapshot of DeepSeek V4.1 Flash, served by DeepSeek under the deepseek-flash API model name.
Glimpse
Elorian
Glimpse is a multimodal visual reasoning model from Elorian, built for visual understanding and complex visual reasoning in the physical world.
Muse Spark Open
Meta
Muse Spark Open is a coding-optimized open-weight model.
Nemotron 3.5 Super VL
NVIDIA
NVIDIA Nemotron 3.5 Super VL is an open multimodal mixture-of-experts model from NVIDIA, with 12B active parameters out of 120B total.
Gemini 4 Argon
Gemini 4 Argon is Google's flagship reasoning model, built for complex software engineering, deep mathematical and scientific problem solving, and autonomous multi-step agents.
Muse Wonder
Meta
Muse Wonder is a frontier reasoning model from Meta, designed for complex personal and professional agentic tasks.
More Step 3.5 Flash comparisons
DeepSeek V4.1 Flash 1017
DeepSeek
DeepSeek V4.1 Flash 1017 is the October 17, 2026 snapshot of DeepSeek V4.1 Flash, served by DeepSeek under the deepseek-flash API model name.
Glimpse
Elorian
Glimpse is a multimodal visual reasoning model from Elorian, built for visual understanding and complex visual reasoning in the physical world.
Beam 1.0 501B A23B
Reflection
Beam 1.0 is Reflection's open-weight reasoning model for coding and long-horizon agentic workflows.
Mercury 3
Inception
Mercury 3 is a diffusion-based language model from Inception, and the lab's most capable model to date.
Moondream 3.1
Moondream
Moondream 3.1 is a vision-language model from Moondream, the successor to Moondream 3 Preview, with 9B total parameters and 2B active (mixture-of-experts).
Muse Spark Open
Meta
Muse Spark Open is a coding-optimized open-weight model.