Models / Ripple 65B Preview
Best home rig for Ripple 65B Preview
Quick answer
How much VRAM does Ripple 65B Preview need?
The weights are about 39.8 GB at Q4_K_M (65B params × 4.89 bits ÷ 8, before context/KV cache). Add memory for context (KV cache) and the runtime. The smallest-memory setup with a measured result here, fully in memory, is Apple M3 Ultra (96 GB) at 15.0 tok/s.
What is the fastest measured hardware for Ripple 65B Preview?
RTX 3090: 25.9 tok/s generation (llama.cpp b12167 llama-server (CUDA), AD-IQ2_XXS (AtomicChat GGUF, 20.0 GB)), from https://inferencekey.com/research/ripple-65b-dense-coder-one-card.
Last updated:
Current model65B dense4 measured resultsfastest plain decode: 25.9 tok/s model info ↗
Will it fit my GPU? Best value GPU per euro →
Real, measured generation speeds for Ripple 65B Preview on home hardware, fastest first. Different quants/engines aren't directly comparable.
| Build | Quant | Engine | tok/s | Prompt processing (tok/s) | Cost | Source |
|---|---|---|---|---|---|---|
| RTX 3090 24 GB: Ripple-65B Preview 2-bit AD-IQ2_XXS on llama.cpp at 23.6-30.9 tok/s (16K ctx, one wrong test in three coding answers) | AD-IQ2_XXS (AtomicChat GGUF, 20.0 GB) | llama.cpp b12167 llama-server (CUDA) | 25.9 | – | inferencekey.com ↗ | |
| Mac Studio M3 Ultra 96 GB: Ripple-65B Preview Q4_K_M on llama.cpp at 15.0 tok/s (13.5 after a 32K prompt, 9.45 at 128K) | Q4_K_M (self-quantized, 38.8 GB) | llama.cpp (Metal) | 15.0 | 126 @512 · est. TTFT @8k ≈65.1 s* | leanzero.net ↗ | |
| RTX A6000 48 GB: Ripple-65B Preview (dense 65B coder) Q4_K_M on llama.cpp at 14.3-16.3 tok/s, prompts up to 256K on one card | Q4_K_M (AtomicChat GGUF, 38.8 GB) | llama.cpp b12167 llama-server (CUDA) | 14.4 | 540 @25,930 · est. TTFT @8k ≈15.2 s* | inferencekey.com ↗ | |
| 4x Intel Arc Pro B70 (Xeon workstation): Ripple-65B Preview Q4_K_M on llama.cpp SYCL at 13.0 tok/s, layer split, 200K context | Q4_K_M (AtomicChat GGUF, 38.8 GB) | llama.cpp b12314 (SYCL), layer split across 4 GPUs | 13.0 | – | intelinside.ai ↗ |
Prompt processing = prefill speed reported by the source, @ prompt length. * est. TTFT @8k is calculated from measured prefill, not measured. Why it matters.
Quality cost of each quant for Ripple 65B Preview
How far each quantization drifts from the full-precision (BF16) model, as published by the quant maker. Lower KL divergence (KLD) is closer to the original; same-top-token % is how often the quant picks the same next token as BF16. Q4_K_M is usually the sweet spot; below Q4 the drift grows fast, which is why RigBase only lists Q4-or-higher speed results.
| Quant | File size | Mean KLD vs BF16 | Same top token |
|---|---|---|---|
| Q8_0 | 68.8 GB | 0.001 | 99.0% |
| AD-Q6_K | 57.7 GB | 0.0016 | 98.7% |
| AD-Q5_K_M | 47.9 GB | 0.0067 | 97.3% |
| AD-Q4_K_M | 40.4 GB | 0.0136 | 95.9% |
| AD-IQ4_XS | 38.7 GB | 0.0149 | 95.6% |
| Q4_K_M (plain llama.cpp, no imatrix) | 38.8 GB | 0.038 | 93.0% |
| AD-IQ3_S below Q4 | 32.4 GB | 0.0327 | 93.4% |
| AD-IQ3_XXS below Q4 | 28.0 GB | 0.0584 | 91.2% |
| AD-IQ2_S below Q4 | 25.3 GB | 0.0959 | 88.5% |
| AD-IQ2_XS below Q4 | 22.1 GB | 0.1778 | 84.2% |
| AD-IQ2_XXS below Q4 | 20.0 GB | 0.2605 | 80.8% |
| AD-IQ1_M below Q4 | 18.5 GB | 0.3743 | 76.8% |
Source: huggingface.co ↗. AtomicChat (quant publisher) GGUF card: llama-perplexity on their 30-language neutral corpus (87 chunks of 4,096 tokens), llama.cpp 9e2a04a with the Ripple port, vs the BF16 original. AD files are AtomicChat's imatrix per-tensor layouts; the plain Q4_K_M is a no-imatrix baseline. KLD from different publishers, datasets or llama.cpp builds isn't directly comparable.
Memory math: how much VRAM / RAM does Ripple 65B Preview need?
Ripple 65B Preview has 65B parameters (source: Ricursive model card: dense 65B, 80 layers (64 sliding-window + 16 global), Apache 2.0, released 4 Nov 2026). Weight size = parameters × bits-per-weight ÷ 8, before context (KV cache) and runtime overhead.
| Format | Bits/weight | Weights | Smallest common memory size that holds the weights alone |
|---|---|---|---|
| Q8_0 | 8.50 | 69.1 GB | 96 GB |
| Q6_K | 6.56 | 53.3 GB | 64 GB |
| Q5_K_M | 5.70 | 46.3 GB | 48 GB |
| Q4_K_M | 4.89 | 39.8 GB | 48 GB |
| Q3_K_M | 4.00 | 32.5 GB | 48 GB |
| Q2_K | 3.16 | 25.7 GB | 32 GB |
Bits-per-weight for GGUF quants are llama.cpp's own figures (llama.cpp quantize README). "As measured" rows apply the source's stated bits-per-weight to the parameter count; real files differ slightly because some layers stay at higher precision. Try other combinations in the VRAM calculator.
Measured on setups with less GPU memory than the Q4_K_M weights (so part of the model lived in system RAM or was streamed from disk):
- RTX 3090 (24 GB): 25.9 tok/s (AD-IQ2_XXS (AtomicChat GGUF, 20.0 GB))
FAQ
How much VRAM does Ripple 65B Preview need?
The weights are about 39.8 GB at Q4_K_M (65B params × 4.89 bits ÷ 8, before context/KV cache). Add memory for context (KV cache) and the runtime. The smallest-memory setup with a measured result here, fully in memory, is Apple M3 Ultra (96 GB) at 15.0 tok/s.
What is the fastest measured hardware for Ripple 65B Preview?
RTX 3090: 25.9 tok/s generation (llama.cpp b12167 llama-server (CUDA), AD-IQ2_XXS (AtomicChat GGUF, 20.0 GB)), from https://inferencekey.com/research/ripple-65b-dense-coder-one-card.
Which GPUs have measured results for Ripple 65B Preview?
4x Intel Arc Pro B70, Apple M3 Ultra, RTX 3090.
Can it run on… 4x Intel Arc Pro B70 · Apple M3 Ultra · RTX 3090
Related: what GPU do you need for 2026 models · all builds