GPT-oss 20B runs at 141 tok/s (140.5 as recorded) on a 128 GB M5 Max, a community run reported by omlx.ai (MLX, MXFP4, 1k context). It needs about 13 GB of unified memory. It is OpenAI's 21B mixture-of-experts model (3.6B active), released under the Apache 2.0 licence.
Want to run GPT-oss 20B faster — or step up to bigger models? Find your Mac →This is a community run: a real run, reported by its source and linked below. The index records it as given and never adjusts it.
| Metric | Value |
|---|---|
| Tokens per second | community141 tok/s |
| Provenance | community run · omlx.ai |
| Quantization | MXFP4 |
| Engine | MLX |
| Time to first token | 0.587s |
| Prompt processing (prefill) | community4774 tok/s at 4K context |
| Context | 1k |
| Figure dated | 2026-10-31 |
| Minimum RAM | 13 GB |
| Parameters | 21B MoE (3.6B active) |
| Chip bandwidth | 614 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 (13 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 |
|---|---|---|---|---|---|---|
| sourced156 tok/s | MXFP4 | llama.cpp | 128 GB | random-token prompt, empty context | vsesmart.cz | Vse Smart review: MacBook Pro M5 Max (40-core GPU, 128 GB), macOS 27.2, High Power mode, on AC; llama.cpp build 11669 (Metal), mean of 3. OpenAI's native MXFP4 release. |
| sourced142 tok/s | MXFP4 | MLX | 128 GB | random-token prompt, empty context | vsesmart.cz | Vse Smart review: MacBook Pro M5 Max (40-core GPU, 128 GB), macOS 27.2, High Power mode, on AC; mlx-lm 0.32.2 (MLX 0.32.4), median of 3. OpenAI's native MXFP4 release. |
| community140.5 tok/s | MXFP4 | MLX | 128 GB | 1k | omlx.ai | 40-core GPU. oMLX v0.7.3, MTP/DFlash off; gpt-oss-20b-MXFP4-Q8 (MXFP4 experts, 8-bit elsewhere); 11.8 GB peak. 137.1 tok/s at 4k. |
| sourced107.51 tok/s | quantization not stated | Ollama | 128 GB | — | tbreak.com | Mac Studio, 40-core GPU. tbreak review: Ollama (version not stated); quantization not stated. Ollama's gpt-oss:20b is the MXFP4 build OpenAI released, so where this run scales the model's other chips it is costed at MXFP4. |
| estimated140 tok/s | MXFP4 | MLX | 64 GB | — | LLMCheck estimator | — |
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) | @4K | @32K | @64K | Prefill | TTFT |
|---|---|---|---|---|---|---|
| llama.cpp · MXFP4 · 128 GB Mac · vsesmart.cz | sourced156 tok/s | — | sourced106 tok/s | — | sourced3909 tok/s | — |
| MLX · MXFP4 · 128 GB Mac · vsesmart.cz | sourced142 tok/s | — | sourced105 tok/s | — | sourced6409 tok/s | — |
| MLX · MXFP4 · 128 GB Mac · omlx.ai | community140.5 tok/s | community137.1 tok/s | community106 tok/s | community84 tok/s | community4774 tok/s | community0.587 s |
Ollama — install it (brew install ollama or the app from ollama.com), then pull the model by its tag, checked on ollama.com on 2026-11-14:
MLX — pick a build from Hugging Face’s MLX listings, then run it with its repo id:
Downloads: Ollama page → · Hugging Face model card →
The 141 tok/s figure above came from MLX at MXFP4; other runtimes and quants run at different speeds.
Prefer a GUI? LM Studio downloads GGUF and MLX builds in one click. Step by step: install Ollama on a Mac.
OpenAI's Apache 2.0 open-weight MoE (Aug 2025): 21B total, 3.6B active, shipped in MXFP4 so it runs in 16 GB. Model card: SWE-bench Verified 53.2, GPQA Diamond 58.6 at medium reasoning. tbreak's review measured 107.5 tok/s on an M5 Max in Ollama, and a community LM Studio run measured ~45 tok/s on a 32 GB Mac mini M6.
Apache 2.0 licence · 21B MoE · 128K context · LLMCheck capability 22/50 · see its leaderboard rank →
Estimated, not measured: the index’s minimum for GPT-oss 20B (13 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 + 12 sliding-window layers (window 128). 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 M5 Max Mac, which runs macOS 26 or later): “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).
GPT-oss 20B needs about 13 GB of unified memory. These Apple Silicon Macs have enough to run it.
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 GPT-oss 20B stacks up against other models on your specific Mac hardware.
Open Compare Tool Full Leaderboard