{"slug": "step-flash", "name": "Step Flash", "category": "ai", "category_name": "AI models and assistants", "description": "", "recommenders": 3, "commenters": 5, "warners": 1, "score": 1.99, "rising": -0.98, "first": "2026-02-19", "latest": "2026-08-03", "why": "Step 3.7 Flash at IQ4_XS beats Qwen 3.6 27B despite the latter having MTP, on all of my actual coding task evaluations in real codebases.", "why_is_quote": true, "has_verdict": false, "sentiment": {"recommend": 2, "neutral": 2, "compare": 0, "warn": 1}, "spark": [0, 0, 0, 1, 0, 0, 1, 0, 0, 0, 0, 0], "listed": true, "loved": [], "caveat": "", "short_quote": "Step 3.7 Flash at IQ4_XS beats Qwen 3.6 27B despite the latter having MTP, on all of my actual coding task evaluations in real codebases.", "icon": false, "domain": null, "github_repo": null, "official_url": "", "pricing": "", "aliases": ["Step 3.7 Flash", "Step 3.5 Flash", "Step-3.5-Flash"], "verdict": null, "praise": [{"label": "", "summary": "", "quotes": [{"id": 778665712, "item_id": 48666607, "url": "https://news.ycombinator.com/item?id=48666607", "by": "girvo", "date": "2026-06-24", "stance": "recommend", "quote": "I use Step 3.7 Flash locally, models are good enough for so many coding tasks even at the lower end!", "excerpt": "Yes, but that was because I could see the writing on the wall with respect to hardware prices being cooked by AI demand, so I built the best computer possible at the time knowing it'd probably need to last me the next 5+ years So not really comparable. I use Step 3.7 Flash locally, models are good enough for so many coding tasks even at the lower end! (Though I note that calling a 200B model \"lower end\" is kind of am", "context": "", "use_case": "local coding model", "switched_from": "", "story": "GLM-5.2 is the step change for open agents"}, {"id": 777008592, "item_id": 48563037, "url": "https://news.ycombinator.com/item?id=48563037", "by": "girvo", "date": "2026-06-16", "stance": "recommend", "quote": "I'm still running Step 3.7 Flash at ~40tk/s decode, 1000tk/s+ prefill on mine and its both very capable and fast enough", "excerpt": "I love my Spark-alike, but they really aren't inference boxes IMO. They're experimentation boxes. A couple of 3080 20GB's for cheap from China, a 5090, an RTX Pro 6000 if you can swing the horrible cost: those are better choices IMO That said, I'm still running Step 3.7 Flash at ~40tk/s decode, 1000tk/s+ prefill on mine and its both very capable and fast enough I got Gemma 31b to run on this at ~22tk/s decode at FP8 ", "context": "", "use_case": "local inference", "switched_from": "", "story": "Running local models is good now"}, {"id": 776768720, "item_id": 48548045, "url": "https://news.ycombinator.com/item?id=48548045", "by": "girvo", "date": "2026-06-15", "stance": "recommend", "quote": "Step 3.7 Flash at IQ4_XS beats Qwen 3.6 27B despite the latter having MTP, on all of my actual coding task evaluations in real codebases.", "excerpt": "Right. Tokens/s decode isn't the most important thing to me: wall clock time for task completion is. And tracking all of that, on my GB10-based Asus box, Step 3.7 Flash at IQ4_XS beats Qwen 3.6 27B despite the latter having MTP, on all of my actual coding task evaluations in real codebases. Qwen seems better at one-shotting things based on vague prompts to an acceptable degree, but thats literally not what I use thes", "context": "runs a GB10-based Asus box", "use_case": "local coding tasks", "switched_from": "", "story": "Ask HN: Has anyone replaced Claude/GPT with a local model for daily coding?"}, {"id": 773211377, "item_id": 48325711, "url": "https://news.ycombinator.com/item?id=48325711", "by": "alfiedotwtf", "date": "2026-05-29", "stance": "recommend", "quote": "So happy 3.7 came out!", "excerpt": "Stepfun 3.5 was my daily driver, and I\u2019ve been blown away at how good it is (even vs Qwen3.6 27B). So happy 3.7 came out!", "context": "", "use_case": "", "switched_from": "Step 3.5", "story": "Step 3.7 Flash"}, {"id": 753168432, "item_id": 47073027, "url": "https://news.ycombinator.com/item?id=47073027", "by": "tallesborges92", "date": "2026-02-19", "stance": "recommend", "quote": "I\u2019ve been using this model for a while, and it\u2019s very fast.", "excerpt": "I\u2019ve been using this model for a while, and it\u2019s very fast. It spent some time thinking but does fewer calls. For example, yesterday I asked the agent to find the Gemini quota limit for their API, and it took 27 seconds and just 2 calls, Opus 4.6 took 33 seconds, but 5 calls with less thinking", "context": "", "use_case": "agentic tasks", "switched_from": "", "story": "Step 3.5 Flash: Fast Enough to Think. Reliable Enough to Act"}]}], "caveats": [{"label": "", "summary": "", "quotes": [{"id": 786552481, "item_id": 49159530, "url": "https://news.ycombinator.com/item?id=49159530", "by": "Swapnoneel", "date": "2026-08-03", "stance": "warn", "quote": "it hallucinates during a lot of tasks, and the output quality isn\u2019t something to write home about", "excerpt": "So, I recently started using the Hermes agent and hooked it up with Step 3.5 Flash. It\u2019s a cheap and fast model, so I thought it would be enough for my workflow. However, I soon found that it hallucinates during a lot of tasks, and the output quality isn\u2019t something to write home about. On the other hand, using an expensive or frontier-level model means I\u2019ll burn through cash fast. I went down the rabbit hole to find", "context": "", "use_case": "cheap model for busywork", "switched_from": "", "story": "Is Hermes Agent unusable without a model router?"}]}], "who": [{"by": "girvo", "context": "runs a GB10-based Asus box", "url": "https://news.ycombinator.com/item?id=48548045", "date": "2026-06-15"}, {"by": "sosodev", "context": "running local models on a 395+ board", "url": "https://news.ycombinator.com/item?id=47077697", "date": "2026-02-19"}], "tags": [], "use_cases": ["cheap model for busywork", "local coding model", "local inference", "local coding tasks", "local llm", "local agentic coding"], "compared": [{"slug": "qwen-3-6", "name": "Qwen 3.6", "co": 2, "switched_from": 0, "switched_to": 0}, {"slug": "minimax", "name": "MiniMax", "co": 2, "switched_from": 0, "switched_to": 0}], "makers": [], "threads": [{"url": "https://news.ycombinator.com/item?id=49159529", "title": "Is Hermes Agent unusable without a model router?", "mentions": 1}, {"url": "https://news.ycombinator.com/item?id=48639840", "title": "GLM-5.2 is the step change for open agents", "mentions": 1}, {"url": "https://news.ycombinator.com/item?id=48555993", "title": "Running local models is good now", "mentions": 1}, {"url": "https://news.ycombinator.com/item?id=48542100", "title": "Ask HN: Has anyone replaced Claude/GPT with a local model for daily coding?", "mentions": 1}, {"url": "https://news.ycombinator.com/item?id=48515454", "title": "RTX 5080 and RTX 3090 Setup: 80 Tok/s on Qwen 3.6 27B Q8", "mentions": 1}, {"url": "https://news.ycombinator.com/item?id=48322451", "title": "Step 3.7 Flash", "mentions": 1}], "sentiment_by_month": {"2025-11": {}, "2025-12": {}, "2026-01": {}, "2026-02": {"recommend": 1, "neutral": 1}, "2026-03": {}, "2026-04": {}, "2026-05": {"recommend": 1}, "2026-06": {"neutral": 1}, "2026-07": {}, "2026-08": {"warn": 1}, "2026-09": {}, "2026-10": {}}, "quotes": [{"id": 786552481, "item_id": 49159530, "url": "https://news.ycombinator.com/item?id=49159530", "by": "Swapnoneel", "date": "2026-08-03", "stance": "warn", "quote": "it hallucinates during a lot of tasks, and the output quality isn\u2019t something to write home about", "excerpt": "So, I recently started using the Hermes agent and hooked it up with Step 3.5 Flash. It\u2019s a cheap and fast model, so I thought it would be enough for my workflow. However, I soon found that it hallucinates during a lot of tasks, and the output quality isn\u2019t something to write home about. On the other hand, using an expensive or frontier-level model means I\u2019ll burn through cash fast. I went down the rabbit hole to find", "context": "", "use_case": "cheap model for busywork", "switched_from": "", "story": "Is Hermes Agent unusable without a model router?"}, {"id": 778665712, "item_id": 48666607, "url": "https://news.ycombinator.com/item?id=48666607", "by": "girvo", "date": "2026-06-24", "stance": "recommend", "quote": "I use Step 3.7 Flash locally, models are good enough for so many coding tasks even at the lower end!", "excerpt": "Yes, but that was because I could see the writing on the wall with respect to hardware prices being cooked by AI demand, so I built the best computer possible at the time knowing it'd probably need to last me the next 5+ years So not really comparable. I use Step 3.7 Flash locally, models are good enough for so many coding tasks even at the lower end! (Though I note that calling a 200B model \"lower end\" is kind of am", "context": "", "use_case": "local coding model", "switched_from": "", "story": "GLM-5.2 is the step change for open agents"}, {"id": 777008592, "item_id": 48563037, "url": "https://news.ycombinator.com/item?id=48563037", "by": "girvo", "date": "2026-06-16", "stance": "recommend", "quote": "I'm still running Step 3.7 Flash at ~40tk/s decode, 1000tk/s+ prefill on mine and its both very capable and fast enough", "excerpt": "I love my Spark-alike, but they really aren't inference boxes IMO. They're experimentation boxes. A couple of 3080 20GB's for cheap from China, a 5090, an RTX Pro 6000 if you can swing the horrible cost: those are better choices IMO That said, I'm still running Step 3.7 Flash at ~40tk/s decode, 1000tk/s+ prefill on mine and its both very capable and fast enough I got Gemma 31b to run on this at ~22tk/s decode at FP8 ", "context": "", "use_case": "local inference", "switched_from": "", "story": "Running local models is good now"}, {"id": 776768720, "item_id": 48548045, "url": "https://news.ycombinator.com/item?id=48548045", "by": "girvo", "date": "2026-06-15", "stance": "recommend", "quote": "Step 3.7 Flash at IQ4_XS beats Qwen 3.6 27B despite the latter having MTP, on all of my actual coding task evaluations in real codebases.", "excerpt": "Right. Tokens/s decode isn't the most important thing to me: wall clock time for task completion is. And tracking all of that, on my GB10-based Asus box, Step 3.7 Flash at IQ4_XS beats Qwen 3.6 27B despite the latter having MTP, on all of my actual coding task evaluations in real codebases. Qwen seems better at one-shotting things based on vague prompts to an acceptable degree, but thats literally not what I use thes", "context": "runs a GB10-based Asus box", "use_case": "local coding tasks", "switched_from": "", "story": "Ask HN: Has anyone replaced Claude/GPT with a local model for daily coding?"}, {"id": 776349970, "item_id": 48521873, "url": "https://news.ycombinator.com/item?id=48521873", "by": "girvo", "date": "2026-06-13", "stance": "neutral", "quote": "my local LLM (Step 3.7 Flash)", "excerpt": "I have a \"task router\" that is a small local LLM on my mac mini (Qwen 3.5 0.8B) that I use to decide (when activated) with Pi whether to route a given task to my local LLM (Step 3.7 Flash) or to <given cloud provider>, if that counts? It works surprisingly well really. Though some of the cloud providers are getting so good and so cheap (GLM 5.1/5.2, MiniMax M3, among others) that the need to use my local one becomes ", "context": "", "use_case": "local LLM", "switched_from": "", "story": "RTX 5080 and RTX 3090 Setup: 80 Tok/s on Qwen 3.6 27B Q8"}, {"id": 773211377, "item_id": 48325711, "url": "https://news.ycombinator.com/item?id=48325711", "by": "alfiedotwtf", "date": "2026-05-29", "stance": "recommend", "quote": "So happy 3.7 came out!", "excerpt": "Stepfun 3.5 was my daily driver, and I\u2019ve been blown away at how good it is (even vs Qwen3.6 27B). So happy 3.7 came out!", "context": "", "use_case": "", "switched_from": "Step 3.5", "story": "Step 3.7 Flash"}, {"id": 753243155, "item_id": 47077697, "url": "https://news.ycombinator.com/item?id=47077697", "by": "sosodev", "date": "2026-02-19", "stance": "neutral", "quote": "I am curious about https://huggingface.co/stepfun-ai/Step-3.5-Flash given that it does parallel token generation.", "excerpt": "Thanks for the additional info. I suspected that MiniMax M2.5 might be a bit too much for this board. 230B-A10B is just a lot to ask of the 395+ even with aggressive quantization. Particularly when you consider that the model is going to spend a lot of tokens thinking and that will eat into the comparatively smaller context window. I switched from the Unsloth 4-bit quant of Qwen3 Coder Next to the official 4-bit quan", "context": "running local models on a 395+ board", "use_case": "local agentic coding", "switched_from": "", "story": "Why AI Velocity Is Becoming a Debt Accelerator"}, {"id": 753168432, "item_id": 47073027, "url": "https://news.ycombinator.com/item?id=47073027", "by": "tallesborges92", "date": "2026-02-19", "stance": "recommend", "quote": "I\u2019ve been using this model for a while, and it\u2019s very fast.", "excerpt": "I\u2019ve been using this model for a while, and it\u2019s very fast. It spent some time thinking but does fewer calls. For example, yesterday I asked the agent to find the Gemini quota limit for their API, and it took 27 seconds and just 2 calls, Opus 4.6 took 33 seconds, but 5 calls with less thinking", "context": "", "use_case": "agentic tasks", "switched_from": "", "story": "Step 3.5 Flash: Fast Enough to Think. Reliable Enough to Act"}]}