Qwen 3.6-35B-A3B runs at 79 tok/s (79.48 as recorded) on a 48 GB M5 Pro, a community run reported by github.com (Ollama, Q4_K_M, ~85-token prompt, 512-token coding answers, median of 5 at T 0.6). It needs about 24 GB of unified memory. It is Alibaba's 35B mixture-of-experts model (3B active), released under the Apache 2.0 licence.
Want to run Qwen 3.6-35B-A3B 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 | community79 tok/s |
| Provenance | community run · github.com |
| Quantization | Q4_K_M |
| Engine | Ollama |
| Context | ~85-token prompt, 512-token coding answers, median of 5 at T 0.6 |
| Figure dated | 2026-10-19 |
| Minimum RAM | 24 GB |
| Parameters | 35B MoE (3B active) |
| Chip bandwidth | 307 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 (24 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 |
|---|---|---|---|---|---|---|
| community95.62 tok/s | 4-bit | LM Studio | 48 GB | ~85-token prompt, 512-token coding answers, median of 5 at T 0.6 | github.com | MacBook Pro, 20-core GPU. LM Studio MLX runtime 1.13.1, lmstudio-community MLX 4-bit; greedy 97.85; flagged for 511 of 512 completion tokens. Run by the Kapitsa engine's developers with their own harness. |
| community94.3 tok/s | 4-bit | MLX | 48 GB | ~85-token prompt, 512-token coding answers, median of 5 at T 0.6 | github.com | MacBook Pro, 20-core GPU. mlx-vlm server 0.7.8, no speculation; greedy 97.43. Only the session's cached 16K follow-up phase was invalid; the decode phase is valid. Run by the Kapitsa engine's developers with their own harness. |
| community81.96 tok/s | Q4_K_M | llama.cpp | 48 GB | ~85-token prompt, 512-token coding answers, median of 5 at T 0.6 | github.com | MacBook Pro, 20-core GPU. llama.cpp b11669 llama-server on Ollama's qwen3.6:35b-a3b-mtp-q4_K_M GGUF with drafting off (MTP n-max 2 reached 107.3, a different workload); greedy 82.19. Run by the Kapitsa engine's developers with their own harness. |
| community79.48 tok/s | Q4_K_M | Ollama | 48 GB | ~85-token prompt, 512-token coding answers, median of 5 at T 0.6 | github.com | MacBook Pro, 20-core GPU. Ollama 0.40.6 on its qwen3.6:35b-a3b-mtp-q4_K_M GGUF with drafting off (draft_num_predict 0; the shipped MTP depth 2 reached 102.4, a different workload); greedy 81.24. Run by the Kapitsa engine's developers with their own harness. |
| community77.8 tok/s | 4-bit | MLX | 48 GB | ~85-token prompt, 512-token coding answers, median of 5 at T 0.6 | github.com | MacBook Pro, 20-core GPU. mlx-lm 0.32.2 (MLX 0.32.4), no MTP path; greedy 83.75. Run by the Kapitsa engine's developers with their own harness. |
| community75.66 tok/s | 4-bit | MLX | 48 GB | ~85-token prompt, 512-token coding answers, median of 5 at T 0.6 | github.com | MacBook Pro, 20-core GPU. Rapid-MLX 0.16.3 (mlx 0.32.4, mlx-lm 0.31.3), MLX 4-bit, no speculation (its catalog turns it off for this MoE); greedy 77.64. Run last in the session (15:27). Run by the Kapitsa engine's developers with their own harness. |
| sourced68.7 tok/s | 4-bit | llama.cpp | 48 GB | single stream; prompt length not stated | frandroid.com | Frandroid's RTX Spark review: MacBook Pro 14 M5 Pro (18-core CPU, 48 GB), llama.cpp b11582 on Metal (llama-bench), a 22 GB 4-bit GGUF whose type the review does not name. |
Ollama — install it (brew install ollama or the app from ollama.com), then pull the model by its tag, checked on ollama.com on 2026-10-24:
MLX — pick a build from Hugging Face’s MLX listings, then run it with its repo id:
Downloads: Search the Ollama library → · model card →
The 79 tok/s figure above came from Ollama at Q4_K_M; 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.
| Chip | Speed | Quant | Engine | Bandwidth |
|---|---|---|---|---|
| M5 Ultra | community153 tok/s | 4-bit | MLX | 1228 GB/s |
| M5 Max | sourced118 tok/s | Q4_K_M | llama.cpp | 614 GB/s |
| M3 Ultra | community106 tok/s | 4-bit | MLX | 819 GB/s |
| M4 Max | estimated105 tok/s | Q4_K_M | MLX | 546 GB/s |
| M1 Ultra | estimated105 tok/s | Q4_K_M | MLX | 800 GB/s |
| M2 Max | estimated92 tok/s | Q4_K_M | MLX | 400 GB/s |
| M4 Pro | estimated74 tok/s | Q4_K_M | Ollama | 273 GB/s |
| M6 | community64 tok/s | 4-bit | MLX | 170 GB/s |
Alibaba's Apache 2.0 35B MoE with 3B active, the most-run local coding model of 2026 (73.4% SWE-bench Verified). About 20 GB at 4-bit, so it fits a 24 GB Mac. On an M5 Max, Liquid AI's published baseline lists 118 tok/s (llama.cpp Q4_K_M) and a community oMLX run measured ~126 (4-bit, MTP off); a 32 GB Mac mini M6 runs it at ~64 (community).
Apache 2.0 licence · 35B MoE · 262K context · LLMCheck capability 38/50 · see its leaderboard rank →
Qwen 3.6-35B-A3B needs about 24 GB of unified memory. These current Apple Silicon Macs have the headroom to run it comfortably:
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 Qwen 3.6-35B-A3B stacks up against other models on your specific Mac hardware.
Open Compare Tool Full Leaderboard