AI models designed for software development have evolved rapidly with significant improvements in capabilities over time. Enhanced reasoning skills, broader context comprehension, and higher task accuracy have made these tools increasingly dependable for daily programming tasks. However, behind these technical advancements comes a tangible cost, both literally and in terms of accessibility. Top-tier frontier models often carry steep pricing structures, especially when deployed heavily for repetitive engineering workflows like large-scale code refactoring.
This reality has prompted many developers, including myself, to search for more affordable alternatives without sacrificing output quality. This does not mean expensive models are not worth their price, but for specific tasks such as code modularization that are repetitive and executed repeatedly, costs can snowball quickly if relying constantly on frontier-class models. This is where finding a free yet capable model becomes highly relevant.
In this search, Step 3.7 Flash by StepFun emerged as an intriguing candidate to test. Built on a Mixture-of-Experts (MoE) architecture, this model features 198 billion total parameters with only around 11 billion active parameters per token, making it as computationally efficient as an 11B dense model while retaining the knowledge capacity of a much larger system.
Interestingly, Step 3.7 Flash’s performance across various benchmarks aligns closely with hands-on real-world experience, particularly for code modularization needs. This article covers its benchmark data, compares real-world usage against competitors in its class like Gemini 3.5 Flash and DeepSeek V4 Flash, and evaluates whether this model deserves a spot as your free go-to option for refactoring.
Getting to Know the Architecture of Step 3.7 Flash
Before diving into performance comparisons, it helps to understand the underlying architecture, which explains how this model achieves high efficiency despite its massive knowledge base.
-
High efficiency MoE architecture
This model combines a 196 billion parameter language backbone with a 1.8 billion parameter visual encoder for native image understanding. Despite its large total parameter count, only about 11 billion parameters are active per token, making its computation much lighter than its total size suggests.
-
Expansive context window
Supports a context window of up to 262,144 tokens, generous enough to handle large codebases all at once.
-
Competitive pricing
Priced at $0.20 per million input tokens and $1.15 per million output tokens, making it one of the most affordable options in its class.
-
Three reasoning levels
Users can select low, medium, or high reasoning levels depending on their needs, helping strike a balance between speed, cost, and analytical depth.
Performance Across Various Benchmarks
Step 3.7 Flash delivers impressive results when benchmarked against both peer models and top-tier frontier models.
Compared to Fellow Flash Models
On the ClawEval-1.1 benchmark measuring agent reliability, Step 3.7 Flash leads with a score of 67.1, comfortably outperforming the nearest competitor at 59.8. On SWE-Bench PRO, it edges out DeepSeek V4 Flash with a score of 56.3 compared to 55.6, while holding a decisive lead on ClawEval-1.1 with 67.1 versus 43.6. All of this is accomplished at a slightly lower cost per token than DeepSeek V4 Flash.
For tasks involving search and tool retrieval, its HLE with Tools score reaches 47.20 percent, a notable jump from its predecessor (Step 3.5 Flash at 35.68 percent), while outperforming Flash models from both DeepSeek V4 and Gemini 3.5.
Compared to Frontier Class Models
Even more impressive is Step 3.7 Flash’s ability to compete with much larger and costlier frontier models.
On SWE-Bench PRO, the model scores 56.3 to take second place overall, trailing only Claude 4 Opus at 64.3. On BrowseComp, its score reaches 75.82 percent, approaching the performance of Claude Opus 4.7 and GLM 5.1. Meanwhile, on DeepSearchQA, it achieves an F1 score of 92.82 percent, comparable to Kimi K2.6, a massive 1-trillion parameter model.
On ResearchRubrics, Step 3.7 Flash actually outperforms GPT 5.5 (71.68 percent compared to 61.50 percent) and comes close to Claude Opus 4.7 at 73.92 percent.
Areas That Still Need Improvement
Despite excelling in many areas, one documented weakness lies in pure reasoning outside the context of tool use. The model achieves roughly 50 percent accuracy in that area, placing it in the 27th percentile. This indicates that Step 3.7 Flash’s primary strengths skew toward agentic execution rather than standalone abstract reasoning.
This likely explains its moderate score on the Artificial Analysis Intelligence Index, as that benchmark measures a combined index of reasoning, general knowledge, math, and coding, rather than pure coding or agentic capability alone.
Hands On Experience, Comparison with Gemini 3.5 Flash and DeepSeek V4 Flash
While benchmark metrics offer essential reference points, real-world testing provides a clearer picture. All three models were tested on an identical task, refactoring a monolithic JavaScript codebase into small modular components, to reveal their true traits beyond the benchmark numbers.
Gemini 3.5 Flash, Smart Yet Rigid in Conversation
Gemini 3.5 Flash resembles an individual of high intellect who, because of that very sophistication, struggles with ordinary conversational flow and contextual nuance. Several observations stood out during refactoring sessions.
- Its general knowledge is broad and the technical solutions provided are often correct.
- Frequently misunderstands instruction intent, especially when requests are not explained in extreme detail.
- Tends to rush to conclusions regarding problem root causes or solutions before fully confirming the analysis, leading to recommendations that miss the actual context.
- Tool use feels unstable, occasionally failing to invoke required functions or passing incorrect parameters.
- Multi-turn conversation feels rigid, as if context must be re-explained from scratch every turn for the model to follow along.
- Due to these drawbacks, refactoring spends more time on manual corrections than actual task completion.
DeepSeek V4 Flash, Fast Yet Token Heavy
DeepSeek V4 Flash displays solid instruction understanding and reliable tool execution. Here are key notes from its evaluation.
- Responses are quick in every conversational turn, keeping the workflow smooth with minimal lag.
- Good comprehension of instructions, leading to few misunderstandings about requested tasks.
- Tool use capabilities are solid and rarely fail when calling required functions.
- Prone to jumping to conclusions on problem diagnostics before verifying whether the analysis is actually accurate.
- Unfortunately, the model is quite verbose, generating long responses that could easily be concise.
- Because of its verbosity, the context window fills up rapidly, posing a challenge for large codebases or long working sessions.
Step 3.7 Flash, Slightly Slower But Highly To The Point
Compared to DeepSeek V4 Flash, Step 3.7 Flash feels slightly slower per conversational turn. However, looking closer, several characteristics make it far more pleasant for refactoring.
- Answers and outputs are straight to the point without unnecessary filler or explanations.
- Token consumption is significantly more efficient, estimated at using about half the tokens of DeepSeek V4 Flash for similar tasks.
- Although responses are not as fast as DeepSeek V4 Flash, the final outputs require fewer follow-up corrections, making the overall workflow efficient.
- The resulting module structure is clean and consistent, well suited for refactoring tasks requiring careful code separation.
- It only modifies code paths directly affected by refactoring without touching or rewriting unrelated sections.
- Code comments are minimal, leaving clean output free from unneeded annotations.
- Thanks to token efficiency, it handles larger codebases in a single session without hitting context window limits prematurely.
Overall, Step 3.7 Flash’s focused and token-efficient nature makes it the most comfortable model specifically tailored for code modularization among the three.
Great for Refactoring, Not Necessarily for New Features
Based on these practical findings, Step 3.7 Flash proves superior for splitting monolithic code into clean, separate modules. Its token efficiency and direct approach keep refactoring structured and on track.
However, for implementing brand new features requiring broader general knowledge and creative problem solving, general-purpose models like Gemini, GPT, or Claude remain superior. This aligns with Step 3.7 Flash’s main limitation in pure reasoning, a quality needed more when designing new feature architectures than when reorganizing existing code.
Free Access via Kilocode and Privacy Considerations
One of the main reasons Step 3.7 Flash by StepFun is particularly compelling for refactoring is its free availability via Kilocode for a limited time. Looking strictly at official API pricing, Step 3.7 Flash is actually not as cheap as DeepSeek V4 Flash. However, the presence of a free tier on Kilocode turns it into a highly cost-effective option for code modularization tasks.
Naturally, there are considerations to keep in mind when utilizing this free access. Users should remain cautious, as the free tier on Kilocode is likely to store and use conversation logs and code sessions for future model training. This data collection practice is fairly standard across the broader AI ecosystem. While this trade-off is generally acceptable for open-source projects or non-sensitive codebases, users should avoid entering confidential or proprietary code when using free service tiers.
Beyond Kilocode, Step 3.7 Flash remains accessible through official API routes such as the StepFun Open Platform, OpenRouter, and NVIDIA NIM for users requiring enterprise-grade privacy or custom pipeline integrations.
Conclusion
For anyone searching for a free, capable AI model for code refactoring or modularization, Step 3.7 Flash is well worth testing. Benchmark data highlighting strong scores on ClawEval-1.1 and SWE-Bench PRO directly reflect real-world performance with high token efficiency and direct outputs, especially when tidying up existing codebases.
While its official API pricing is not quite as cheap as DeepSeek V4 Flash, the availability of free access on Kilocode makes it a zero-cost option worth exploring. Despite the privacy consideration regarding training log retention, Step 3.7 Flash via Kilocode remains a compelling choice for refactoring code without committing to paid subscriptions.
