I Switched My Local Coding Model to Step 3.5 Flash

Getting Started#

I tend to test new model releases whenever they look especially strong at coding. This time, the model I tried was Step 3.5 Flash.

To be clear, I do not do most of my day-to-day coding with local models. My main workflow is still built around commercial models, and local models are closer to something I test through Cline whenever a notable new release appears.

What I learned from running several models on a Mac Studio M3 Ultra is that speed matters a lot if you want to use an LLM for coding. It starts to feel fairly comfortable above 50 tok/s, and once it drops below 30 tok/s, it becomes frustrating very quickly.

This is not a long benchmark breakdown. I simply want to explain why this model caught my attention, what felt good when I used it as a local coding model, and how far I would actually recommend it.

Why Step 3.5 Flash?#

Before this, I had been splitting my usage between MiniMax M2.1 for coding and GLM 4.7 for more general tasks. Neither model was bad, but for coding work I still wanted outputs that felt a bit more stable and a bit faster.

That is when StepFun's Step 3.5 Flash stood out. According to the official model card, it uses a 196B MoE architecture, activates 11B parameters at runtime, supports a 256K context window, and is released under Apache 2.0. It also showed strong coding numbers such as 74.4% on SWE-bench Verified.

I do not choose models based on benchmark numbers alone. What stood out instead was how stable the generated code felt in testing. On simpler tasks, it honestly felt good enough to compare with Sonnet 4.5.

What Felt Good in Actual Use#

The first thing I liked was that the code felt relatively stable.

Tasks that used to need one or two extra rounds of explanation started finishing with shorter instructions. It felt especially solid on things like structured code, splitting functions cleanly, and keeping types aligned.

The second point was that I liked its language behavior much more.

Among the local coding models I had tried before, MiniMax was the one I preferred most. But it mixed in Chinese characters surprisingly often, and its Korean was also pretty disappointing. Step 3.5 Flash handled Korean much more naturally, and it almost never threw unexpected Chinese characters into the output.

What felt especially unusual was that most of its reasoning stayed in the same language as the input. I honestly cannot remember another model that matched the input language that consistently during reasoning.

The third point was that it felt more usable locally than I expected.

The official announcement mentions high throughput numbers on the API side, but local hardware obviously does not reproduce those numbers directly. In my setup, it runs much slower than that. Even so, for short edits and repeated code generation, it felt less like something I had to tolerate and more like something I could keep running in the background.

It Is Not a Universal Model#

I would not recommend this model for every kind of work.

For broader tasks such as general conversation or creative writing, other models can still be a better fit. Step 3.5 Flash felt more like a model that is clearly strong at a specific kind of work than one that can cover everything by itself.

Managing expectations also matters.

On a Mac in particular, prefill is simply too slow. As the context gets longer, the wait before the first useful response becomes obvious, and at that point it is hard to get anywhere near the productivity of commercial tools, especially a workflow centered on Claude Code.

Another weak point was how many tokens it seemed to spend on reasoning. Even on relatively simple tasks, it sometimes produced longer reasoning than I expected, which made both the perceived speed and the total token cost feel less efficient.

That is why I see this less as a replacement for my main coding setup and more as a model I test through Cline to understand a new release. It works well enough for short coding loops such as writing, editing, or refactoring code, but if you expect it to carry your primary coding workflow, the limit shows up quickly.

Who Is It For?#

I think it is worth trying in cases like these.

• developers looking for a local model focused on coding
• teams that want to keep more privacy with open-weight models
• workflows that need a model for code generation or code edits
• setups where a dedicated coding model is separated from a general-purpose model

If you want one model to handle creative writing, conversation, and long essays as well, this may not match that expectation.

Closing#

Among the local coding models I have tried recently, Step 3.5 Flash left a pretty good impression.

It is not a perfect all-around model, but if the standard is "an open-weight model focused on coding," I think it is easy to recommend.

If you are building a local coding setup and your current model feels a little ambiguous, Step 3.5 Flash is a candidate worth switching to. At least for me, it has become the model I reach for first among recent local coding options.