Run KAT-Coder-V2.5 on M5 Max

KAT-Coder-V2.5 runs at 89 tok/s (89.3 as recorded) on a 128 GB M5 Max, a community run reported by omlx.ai (MLX, OptiQ-4bit, 1k context). It needs about 24 GB of unified memory. It is Kwaipilot's 35B mixture-of-experts model (3B active), released under the Apache 2.0 licence.

Speed
community89
tok/s
First Token
community0.316
seconds
RAM Needed
24
GB minimum
Engine
MLX
of this figure
Want to run KAT-Coder-V2.5 faster — or step up to bigger models? Find your Mac →

The figure

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.

MetricValue
Tokens per secondcommunity89 tok/s
Provenancecommunity run · omlx.ai
QuantizationOptiQ-4bit
EngineMLX
Time to first token0.316s
Prompt processing (prefill)community4018 tok/s at 4K context
Context1k
Figure dated2026-08-07
Minimum RAM24 GB
Parameters35B MoE (3B active)
Chip bandwidth614 GB/s

Every figure for KAT-Coder-V2.5 on the M5 Max

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.

SpeedQuantEngineMac RAMContextSourceNote
community89.3 tok/sOptiQ-4bitMLX128 GB1komlx.ai40-core GPU. oMLX v0.5.7, MTP/DFlash off; mlx-community/KAT-Coder-V2.5-Dev-OptiQ-4bit (mixed 4/8-bit). 84.8 tok/s at 4k.
Provenance key
sourced
Published figure — vendor or press review
estimated
community
Linked individual run — blog, forum or submission

Speed by context depth: KAT-Coder-V2.5 on the M5 Max

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.

RunDecode (row)@4KPrefillTTFT
MLX · OptiQ-4bit · 128 GB Mac · omlx.aicommunity89.3 tok/s
1k
community84.8 tok/scommunity4018 tok/s
at 4K
community0.316 s

Setup: run KAT-Coder-V2.5 on a Mac

Ollama — the index has not verified an Ollama tag for KAT-Coder-V2.5, 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:

pip install -U mlx-lm
mlx_lm.generate --model <repo-id> --prompt "Hello"

Downloads: Search the Ollama library → · Hugging Face model card →

The 89 tok/s figure above came from MLX at OptiQ-4bit; 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.

KAT-Coder-V2.5 on other Apple Silicon chips

ChipSpeedQuantEngineBandwidth
M4 Maxcommunity94 tok/s4-bitMLX546 GB/s
M3 Maxestimated84 tok/sQ4_K_M—400 GB/s
M4 Proestimated66 tok/sQ4_K_MOllama273 GB/s
M6estimated47 tok/sQ4_K_MMLX170 GB/s
M5estimated44 tok/sQ4_K_M—153 GB/s

For the estimated figures here, the bandwidth formula’s typical error for small MoEs (≤40B total) is ±15%, against 88 measured runs of 15 models (how it is computed).

About KAT-Coder-V2.5

Kwaipilot's Apache 2.0 agentic coder; the open-weight 'Dev' release is a 35B MoE with 3B active, built on Qwen3.6-35B-A3B. About 20 GB at 4-bit, so it fits a 24 GB Mac; a community oMLX run measured 89.3 tok/s on an M5 Max (OptiQ 4-bit, MTP off), below the index's earlier 117 estimate.

Apache 2.0 licence · 35B MoE · 256K context · LLMCheck capability 39/50 · see its leaderboard rank →

Memory by context length

Estimated, not measured: the index’s minimum for KAT-Coder-V2.5 (24 GB — the weights at the quant that fits, plus ~2 GB) plus the KV cache for that many tokens, computed from its config.json. hybrid: full attention on 10 of 40 layers; Gated DeltaNet layers keep a constant state. 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).

ContextKV cache (16-bit)RAM neededRAM, q8 KVSmallest tier that fitsWithin the default GPU limit
8K~0.2 GB~24.1 GB~24.0 GB32 GB Macsame
32K~0.7 GB~24.6 GB~24.3 GB32 GB Macsame
64K (coding-agent context)~1.3 GB~25.3 GB~24.6 GB32 GB Mac raise GPU limit36 GB Mac
128K~2.7 GB~26.6 GB~25.3 GB32 GB Mac raise GPU limit36 GB Mac

System requirements

Macs with the M5 Max

🛒 Get a Mac that runs KAT-Coder-V2.5

KAT-Coder-V2.5 needs about 24 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.

Compare More Models

See how KAT-Coder-V2.5 stacks up against other models on your specific Mac hardware.

Open Compare Tool Full Leaderboard