The LLMCheck index estimates Step-3.7-Flash at ~26 tok/s on an M4 Max at Q4_K_M, scaled from its community-measured M5 Max run with the published bandwidth model, not a measured run on this chip. It needs about 65 GB of unified memory. It is StepFun's 198B mixture-of-experts model (11B active), released under the Apache 2.0 licence.
Want to run Step-3.7-Flash faster — or step up to bigger models? Find your Mac →This is an LLMCheck index estimate from its published two-term bandwidth model, not a measured run; the index labels it so everywhere it appears. The bandwidth formula’s typical error for large MoEs is ±34% (mostly below the measured figure), against 70 measured runs of 18 models (how it is computed). Own this Mac? Submit a real run →
| Metric | Value |
|---|---|
| Tokens per second | estimated~26 tok/s |
| Provenance | estimate · how estimates are made |
| Quantization | Q4_K_M |
| Minimum RAM | 65 GB |
| Parameters | 198B MoE (11B active) |
| Chip bandwidth | 546 GB/s |
Each row as the index holds it, with its source. “Mac RAM” is the memory of the Mac a run used, not the model’s requirement (65 GB). The headline figure follows the reference rule: the lower-median measured plain-decode run at 5 bits or fewer (a run whose quant was not stated counts only when no stated-quant run qualifies), else the estimate; a run marked “not used to anchor estimates” is listed here but never becomes the headline.
| Speed | Quant | Engine | Mac RAM | Context | Source | Note |
|---|---|---|---|---|---|---|
| sourced51.2 tok/s | Q4_K_S | llama.cpp | 128 GB | TG128 from an empty prompt | huggingface.co | StepFun's GGUF card, Mac Studio M4 Max 128 GB: llama-batched-bench on the stepfun-ai/llama.cpp step3.7 branch, before upstream support (b9473, 2 Jun), so it does not anchor; 112 GB. 48.51 tok/s at 2K, 39.72 at 32K.; not used to anchor estimates |
| sourced49.58 tok/s | IQ4_XS | llama.cpp | 128 GB | TG128 from an empty prompt | huggingface.co | StepFun's GGUF card, Mac Studio M4 Max 128 GB: llama-batched-bench on the stepfun-ai/llama.cpp step3.7 branch, before upstream support (b9473, 2 Jun), so it does not anchor; 105 GB. 47.78 tok/s at 2K, 38.93 at 32K.; not used to anchor estimates |
Measured figures only, each from its source’s own session; the index does not estimate long-context or prefill speed, so a missing cell is —. Each @ column is plain decode with that many tokens already in the prompt, from the same session as the row’s short-context figure: decode slows as the KV cache it reads grows. Prefill is prompt processing, the tokens per second a runtime reads a prompt at; with long agent contexts it sets how long you wait before the first token.
| Run | Decode (row) | @8K | @16K | @32K | @64K | @128K | Prefill |
|---|---|---|---|---|---|---|---|
| llama.cpp · Q4_K_S · 128 GB Mac · huggingface.co | sourced51.2 tok/s | sourced46.43 tok/s | sourced43.77 tok/s | sourced39.72 tok/s | sourced32.74 tok/s | sourced22.23 tok/s | sourced404 tok/s |
| llama.cpp · IQ4_XS · 128 GB Mac · huggingface.co | sourced49.58 tok/s | sourced45.66 tok/s | sourced43.29 tok/s | sourced38.93 tok/s | sourced32.44 tok/s | sourced22.1 tok/s | sourced411 tok/s |
Ollama — the index has not verified an Ollama tag for Step-3.7-Flash, so it doesn’t print one (a guessed tag fails with “pull model manifest: file does not exist”). Search the Ollama library →
MLX — pick a build from Hugging Face’s MLX listings, then run it with its repo id:
Downloads: Hugging Face repo → · Hugging Face model card →
Prefer a GUI? LM Studio downloads GGUF and MLX builds in one click. Step by step: install Ollama on a Mac.
StepFun's Apache 2.0 198B MoE (11B active, 23 May 2026), image input, 256K context; it supersedes Step-3.5-Flash. Card: SWE-Bench Pro 56.3 (vendor scaffold), Terminal-Bench 2.1 59.5. The oQ2e MLX build is 63 GB and IQ4_XS ~105 GB, so it needs 96–128 GB; upstream llama.cpp has run it since b9473, and community oMLX runs measured ~48–51 tok/s on an M5 Max at 2–3 bits.
Apache 2.0 licence · 198B MoE · 256K context · LLMCheck capability 41/50 · see its leaderboard rank →
Estimated, not measured: the index’s minimum for Step-3.7-Flash (65 GB — the weights at the quant that fits, plus ~2 GB) plus the KV cache for that many tokens, computed from its config.json. 12 full layers + 33 sliding-window layers (window 512). A 16-bit KV cache is the default in llama.cpp, Ollama, LM Studio and mlx-lm; a q8 cache halves the KV part. Coding agents hold 64K tokens or more. Runtimes also add a buffer that grows with context, so leave a few GB spare. macOS lets the GPU use only part of a Mac’s memory by default (figures for an M4 Max Mac on macOS 15, the lower default; macOS 26 allows more): “fits” means within the GPU memory a Mac of that tier can use with the limit raised, a tier tagged raise GPU limit fits only after sudo sysctl iogpu.wired_limit_mb raises it (the tooltip has the exact command), and the last column is the smallest Mac memory size whose default limit holds it (the GPU memory limit).
| Context | KV cache (16-bit) | RAM needed | RAM, q8 KV | Smallest tier that fits | Within the default GPU limit |
|---|---|---|---|---|---|
| 8K | ~0.5 GB | ~65.2 GB | ~65.1 GB | 128 GB Mac | 96 GB Mac |
| 32K | ~1.7 GB | ~66.4 GB | ~65.7 GB | 128 GB Mac | 96 GB Mac |
| 64K (coding-agent context) | ~3.3 GB | ~68.0 GB | ~66.5 GB | 128 GB Mac | 96 GB Mac |
| 128K | ~6.5 GB | ~71.2 GB | ~68.1 GB | 128 GB Mac | 96 GB Mac |
Step-3.7-Flash needs about 65 GB of unified memory. These Apple Silicon Macs have enough to run it. Apple builds the higher-memory configurations to order; Amazon does not stock them, so the used option links Amazon Renewed listings.
As an Amazon Associate, LLMCheck earns from qualifying purchases. These affiliate links cost you nothing extra and help keep the LLMCheck index free.
See how Step-3.7-Flash stacks up against other models on your specific Mac hardware.
Open Compare Tool Full Leaderboard