Nanbeige4.3-33B runs at 25 tok/s (24.5 as recorded) on a 128 GB M5 Max, a community run reported by omlx.ai (MLX, oQ4e, 1k context). It needs about 17 GB of unified memory. It is Nanbeige's 33.1B dense model, released under the Apache 2.0 licence.
Want to run Nanbeige4.3-33B 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 | community25 tok/s |
| Provenance | community run · omlx.ai |
| Quantization | oQ4e |
| Engine | MLX |
| Time to first token | 1.665s |
| Prompt processing (prefill) | community613 tok/s at 4K context |
| Context | 1k |
| Figure dated | 2026-11-03 |
| Minimum RAM | 17 GB |
| Parameters | 33.1B |
| 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 (17 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 |
|---|---|---|---|---|---|---|
| community27.4 tok/s | 4-bit | MLX | 128 GB | 1k | omlx.ai | 40-core GPU. oMLX v0.7.3, MTP/DFlash/SpecPrefill off; Nanbeige4.3-33B-MLX-4bit; 18.2 GB peak. 24.1 tok/s at 4k, 18.3 at 32k. |
| community24.5 tok/s | oQ4e | MLX | 128 GB | 1k | omlx.ai | 40-core GPU. oMLX v0.7.3, MTP/DFlash off; Nanbeige4.3-33B-oQ4e; 18.9 GB peak. 24.3 tok/s at 4k. |
| community15.8 tok/s | oQ8e | MLX | 128 GB | 1k | omlx.ai | 40-core GPU. oMLX v0.7.3, MTP/DFlash off; Nanbeige4.3-33B-oQ8e; 33.6 GB peak. 15.5 tok/s at 4k. |
| estimated25 tok/s | Q4_K_M | llama.cpp | 48 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 | @8K | @16K | @32K | @64K | @128K | Prefill | TTFT |
|---|---|---|---|---|---|---|---|---|---|
| MLX · 4-bit · 128 GB Mac · omlx.ai | community27.4 tok/s | community24.1 tok/s | community21.1 tok/s | community22.4 tok/s | community18.3 tok/s | community13.6 tok/s | community8.9 tok/s | community528 tok/s | community1.273 s |
| MLX · oQ4e · 128 GB Mac · omlx.ai | community24.5 tok/s | community24.3 tok/s | — | — | — | — | — | community613 tok/s | community1.665 s |
| MLX · oQ8e · 128 GB Mac · omlx.ai | community15.8 tok/s | community15.5 tok/s | community15 tok/s | community14.3 tok/s | community12.8 tok/s | community10.7 tok/s | — | community670 tok/s | community1.448 s |
Ollama — the index has not verified an Ollama tag for Nanbeige4.3-33B, 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 →
The 25 tok/s figure above came from MLX at oQ4e; 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.
Nanbeige's Apache 2.0 dense 33.1B model (24 Oct 2026), 512K context. Artificial Analysis ranks it level with Qwen3.8-27B at the top of its small open models. Vendor-run card: Terminal-Bench 2.1 80.9, SWE-Bench Pro 60.3 (mini-swe-agent, not Scale SEAL), GPQA Diamond 89.5. IQ3_XXS is 14.5 GB and Q4_K_M 20.8 GB, so it targets 24 GB Macs; stock llama.cpp runs it (Llama architecture), and community oMLX runs measured ~25–27 tok/s on an M5 Max at 4-bit.
Apache 2.0 licence · 33.1B · 512K context · LLMCheck capability 46/50 · see its leaderboard rank →
Estimated, not measured: the index’s minimum for Nanbeige4.3-33B (17 GB — the weights at the quant that fits, plus ~2 GB) plus the KV cache for that many tokens, computed from its config.json. GQA on all 52 layers (8 kv heads x 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).
Nanbeige4.3-33B needs about 17 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.
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 Nanbeige4.3-33B stacks up against other models on your specific Mac hardware.
Open Compare Tool Full Leaderboard