The LLMCheck index estimates Ripple-65B Preview at ~26 tok/s on an M5 Ultra at Q4_K_M, a figure from its published bandwidth model, not a measured run. It needs about 33 GB of unified memory. It is Ricursive Intelligence's 64.8B dense model, released under the Apache 2.0 licence.
Want to run Ripple-65B Preview 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 dense models is ±9%, against 75 measured runs of 20 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 | 33 GB |
| Parameters | 64.8B |
| Chip bandwidth | 1228 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 (33 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 |
|---|---|---|---|---|---|---|
| community23.3 tok/s | 4-bit | MLX | 256 GB | 1K context, one stream | huggingface.co | Mac Studio M5 Ultra 256 GB, GPU core count not stated. mlx-lm 0.32.2 --trust-remote-code with the repo's own ricursive.py model file (mlx-lm has no Ripple architecture yet), a custom loader, so it does not anchor; 37 GB, 4.5 bpw. 22.0 tok/s at 32K.; not used to anchor estimates |
| community13.6 tok/s | 8-bit | MLX | 96 GB | 1k | omlx.ai | 64-core GPU (1.2 TB/s, the same bandwidth as the 80-core part). oMLX v0.7.3, MTP/DFlash off; Ripple-65B-Preview-MLX-8bit with the repo's own ricursive.py model file (mlx-lm has no Ripple architecture yet), a custom loader, so it does not anchor; 65.1 GB peak. The 1k figure from this session's context table (the linked page is its 4k run, 13.7 tok/s); 13.1 at 32k.; not used to anchor estimates |
| community12.8 tok/s | 8-bit | MLX | 256 GB | 1K context, one stream | huggingface.co | Mac Studio M5 Ultra 256 GB, GPU core count not stated. mlx-lm 0.32.2 --trust-remote-code with the repo's own ricursive.py model file (mlx-lm has no Ripple architecture yet), a custom loader, so it does not anchor; 69 GB, 8.5 bpw. 12.3 tok/s 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) | @4K | @32K | Prefill | TTFT |
|---|---|---|---|---|---|
| MLX · 8-bit · 96 GB Mac · omlx.ai | community13.6 tok/s | community13.7 tok/s | community13.1 tok/s | community455 tok/s | community2.337 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 →
Prefer a GUI? LM Studio downloads GGUF and MLX builds in one click. Step by step: install Ollama on a Mac.
Ricursive Intelligence's Apache 2.0 dense 65B preview (4 Nov 2026), trained from scratch, 1M context. Self-reported card: SWE-bench Verified 76.5 (OpenCode harness), Terminal-Bench 2.1 80.5, GPQA Diamond 85.2. Q3_K_M is 31 GB and Q4_K_M 38.8 GB, so it needs a 48–64 GB Mac; stock llama.cpp b12167+ and Ollama (ripple-preview:65b) run it, while its MLX builds still need a bundled model file (community runs: ~17–23 tok/s on M3/M5 Ultra at 4-bit). Not related to Recursive, the maker of Flint.
Apache 2.0 licence · 64.8B · 1000K context · LLMCheck capability 42/50 · see its leaderboard rank →
Estimated, not measured: the index’s minimum for Ripple-65B Preview (33 GB — the weights at the quant that fits, plus ~2 GB) plus the KV cache for that many tokens, computed from its config.json. 16 full layers + 64 sliding-window layers (window 1024). 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 Ultra 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).
Ripple-65B Preview needs about 33 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 Ripple-65B Preview stacks up against other models on your specific Mac hardware.
Open Compare Tool Full Leaderboard