Ripple-65B-Preview
65BRipple
A dense 65B preview model from ricursiveAI under Apache 2.0, built for agentic coding. Only every fifth layer (16 of 80) is global attention, and its cache is the one that grows: 64 KB per token, so a 32K window adds 1.9 GB and 128K adds about 8.5 GB. The other 64 layers use a 1,024-token sliding window, a fixed 0.4 GB in total, and the window is 2,097,152 tokens. AtomicChat's files named AD are larger than the plain type of the same name (its IQ2_XXS is 20.0 GB, 2.47 bits per weight). IQ1_M needs about 20.0 GB at 4K, comfortable on a 24 GB card. At IQ2_XXS it needs about 21.6 GB at 4K, tight on a 24 GB card. IQ2_S needs about 27.2 GB at 4K, comfortable on a 32 GB card. At Q4_K_M, the plain 38.8 GB file, it needs about 41.2 GB at 4K, comfortable on a 48 GB card. Sizes come from AtomicChat's GGUF files, and no per-level quality loss has been published.
2097K
Max Context
11
Quant Variants
GGUF Q8_0
Best Quality
—
Perplexity rise (best build)
Quantization Variants
Per-quant VRAM, quality loss, and inference speed on RTX 4090
Measured = run on this site · Estimated = not run here: calculated, or a figure this site has not verified · Community = public reports
| Format | Level | BPW | VRAM | PPL Loss | Speed | Source | Actions |
|---|---|---|---|---|---|---|---|
| GGUF | IQ1_M | 2.28 | 20.0 GB | — | — | Estimated | |
| GGUF | IQ2_XXS | 2.47 | 21.6 GB | — | — | Estimated | |
| GGUF | IQ2_XS | 2.73 | 23.8 GB | — | — | Estimated | |
| GGUF | IQ2_S | 3.13 | 27.2 GB | — | — | Estimated | |
| GGUF | IQ3_XXS | 3.45 | 29.9 GB | — | — | Estimated | |
| GGUF | IQ3_S | 4 | 34.5 GB | — | — | Estimated | |
| GGUF | IQ4_XS | 4.78 | 41.1 GB | — | — | Estimated | |
| GGUF | Q4_K_M | 4.79 | 41.2 GB | — | — | Estimated | |
| GGUF | Q5_K_M | 5.92 | 50.8 GB | — | — | Estimated | |
| GGUF | Q6_K | 7.12 | 60.9 GB | — | — | Estimated | |
| GGUF | Q8_0 | 8.5 | 72.6 GB | — | — | Estimated |
Which file each size is for
A file size is not a memory figure: the VRAM column is the calculator's total at 4K context — weights plus KV cache plus a 10% activation buffer. Units differ on purpose: file sizes are decimal GB, as Hugging Face lists them; the VRAM column is GiB (1024³ bytes) with 2% added to the weights, so 7.12 GB of file counts as about 6.8 GiB there. A vision projector (mmproj), where the repo ships one, is a separate download and is in neither number.
- GGUF IQ1_MAtomicChat/Ripple-65B-Preview-GGUF · Ripple-65B-Preview-AD-IQ1_M.gguf · 18.49 GB (18,493,392,352 bytes) · listing read 2026-11-10
- GGUF IQ2_XXSAtomicChat/Ripple-65B-Preview-GGUF · Ripple-65B-Preview-AD-IQ2_XXS.gguf · 20.01 GB (20,007,273,952 bytes) · listing read 2026-11-10
- GGUF IQ2_XSAtomicChat/Ripple-65B-Preview-GGUF · Ripple-65B-Preview-AD-IQ2_XS.gguf · 22.12 GB (22,120,482,272 bytes) · listing read 2026-11-10
- GGUF IQ2_SAtomicChat/Ripple-65B-Preview-GGUF · Ripple-65B-Preview-AD-IQ2_S.gguf · 25.31 GB (25,310,774,752 bytes) · listing read 2026-11-10
- GGUF IQ3_XXSAtomicChat/Ripple-65B-Preview-GGUF · Ripple-65B-Preview-AD-IQ3_XXS.gguf · 27.96 GB (27,964,982,752 bytes) · listing read 2026-11-10
- GGUF IQ3_SAtomicChat/Ripple-65B-Preview-GGUF · Ripple-65B-Preview-AD-IQ3_S.gguf · 32.41 GB (32,411,600,352 bytes) · listing read 2026-11-10
- GGUF IQ4_XSAtomicChat/Ripple-65B-Preview-GGUF · Ripple-65B-Preview-AD-IQ4_XS.gguf · 38.72 GB (38,722,143,712 bytes) · listing read 2026-11-10
- GGUF Q4_K_MAtomicChat/Ripple-65B-Preview-GGUF · Ripple-65B-Preview-Q4_K_M.gguf · 38.80 GB (38,798,820,576 bytes) · listing read 2026-11-10
- GGUF Q5_K_MAtomicChat/Ripple-65B-Preview-GGUF · Ripple-65B-Preview-AD-Q5_K_M.gguf · 47.93 GB (47,929,624,032 bytes) · listing read 2026-11-10
- GGUF Q6_KAtomicChat/Ripple-65B-Preview-GGUF · file name not recorded · 57.67 GB (57,666,225,856 bytes) · 2 parts · listing read 2026-11-10
- GGUF Q8_0AtomicChat/Ripple-65B-Preview-GGUF · file name not recorded · 68.84 GB (68,840,113,568 bytes) · 2 parts · listing read 2026-11-10
Similar models
Compare with GPT-OSS 120BGPT-OSS 120B
OpenAI GPT-OSS
The big GPT-OSS (117B total / 5.1B active). Native MXFP4 checkpoint is ~61GB — fits one 80GB card or a 128GB unified-memory Mac. Partial offload on 24GB consumer cards is slow but works.
GLM-4.5-Air
Zhipu GLM-4.5
Zhipu's agentic/reasoning MoE (106B total / 12B active). Q4 ~64GB — the sweet spot is a 96GB+ unified Mac or 2× 48GB cards. Strong tool-calling for its class.
Qwen3.8-Flash-Next
Qwen3.8
Qwen's preview of its next architecture: a 125B mixture-of-experts model with about 6B parameters active per token. Released in August, it also carries a 51B n-gram lookup table. llama.cpp reads that table lazily from RAM or disk, so it never needs VRAM. It does need disk space and is bundled inside the GGUF files, which is why they are far larger than the sizes below. Only 12 of its 48 layers keep a growing KV cache, so even the full 256K window adds about 7.5 GB. At Q4_K_M it needs about 79 GB at 4K context, comfortable on a 128 GB Mac or DGX Spark; at 256K it needs about 87 GB, still comfortable on the Spark but tight on the Mac, whose GPU gets about three quarters of its memory. A single 24 or 32 GB card runs it only with experts offloaded to system RAM. llama.cpp supports it as `qwen4exp`. Sizes use the calculator's generic rates, since no file size isolates the model from the table, and no per-level quality loss has been published.
Muse Spark Open 816B-A42B
Meta Muse
Meta's largest open-weight model, released in October under the Muse Model License. It is an 816B mixture-of-experts model with about 42B parameters active per token, built for coding and agent work, with image and audio input. It is a single-machine model only at its smallest build. At Q2_K (ggml-org's 304 GB file) it needs about 319 GB at 4K context and still only 327 GB at 128K, because just 16 of its 62 layers keep a growing cache — comfortable on a 512 GB Mac Studio and tight across four 80 GB H100s. The full 1M window brings it to about 389 GB. Q4_K_S (470 GB) and Q8_0 (872 GB) fit no single machine in this index. The vision and audio encoders ship as separate files of 2–4 GB and are not included. Sizes come from the published GGUF files; Meta's benchmark claims are its own and no per-level quality loss has been published. The licence is not a standard open-source one: Meta's repository is gated, so downloads need an accepted access request; products built on it must show "Built with Muse Spark Open"; and a company with over $20M in annual revenue must sign a separate agreement with Meta before offering it as an API service.
Running Ripple-65B-Preview locally
At Q4_K_M and 4K of context, Ripple-65B-Preview needs about 41.2 GB — 36.9 GB of weights, 0.63 GB of KV cache and a 3.8 GB activation buffer. The smallest cards in this index that clear that comfortably are 48 GB ones — 3 of them, from the A40 48G up — and 17 of the 78 cards here do. These are calculated figures, not measurements: the estimate stops counting a card as comfortable at 88% of its VRAM, which is roughly the room a desktop session needs.
What longer context costs
Going from 4K to 32K adds about 1.75 GB, taking the total to 43.1 GB. That is gentler than the parameter count suggests: only 16 of 80 layers keep a KV cache that grows with context, and 64 more cap theirs at a sliding window, so the cache scales at a fraction of the usual rate. The model's native window is 2M; holding all of it at Q4_K_M would need about 182 GB.
Which build to download
This index only tracks Ripple-65B-Preview in GGUF, across 11 levels (IQ1_M, IQ2_XXS, IQ2_XS, IQ2_S, IQ3_XXS, IQ3_S, IQ4_XS, Q4_K_M, Q5_K_M, Q6_K, Q8_0). No per-level perplexity sweep has been published for this model, so the quality column is empty rather than estimated — within one model, more bits per weight is the only ordering the data supports. No throughput has been measured for this model on this site's hardware. GGUF runs on llama.cpp and Ollama across NVIDIA, AMD and Apple silicon; AWQ and GPTQ target vLLM on CUDA and ROCm; EXL2 is ExLlamaV2 — CUDA only, and now archived, so local use rather than serving.
Common questions
- How much VRAM does Ripple-65B-Preview need?
- About 41.2 GB at Q4_K_M with 4K of context and batch 1, rising to roughly 43.1 GB at 32K. That figure is weights plus KV cache plus a 10% activation buffer, calculated from the model's architecture rather than measured on a card.
- Will Ripple-65B-Preview run on a 48GB GPU?
- Yes — at Q4_K_M and 4K context it needs about 41.2 GB, which leaves 6.8 GB spare on a A40 48G, or on any of the other 48 GB cards here (3 in all). That is the smallest size in this index that clears it comfortably; 17 of 78 do.
- Which quantization of Ripple-65B-Preview should I use?
- No published quality comparison exists for this model, so pick by footprint: Q4_K_M is the level most people run, and a higher bits-per-weight level is more faithful. All 11 levels in the index are IQ1_M, IQ2_XXS, IQ2_XS, IQ2_S, IQ3_XXS, IQ3_S, IQ4_XS, Q4_K_M, Q5_K_M, Q6_K, Q8_0.
Where this model fits
Sized at Q4_K_M with a 4K context window, smallest card first. Comfortable means the estimate uses at most 88% of the memory.
- Runs comfortably on
- A40 48G48GB · +6.8L40S 48G48GB · +6.8Radeon PRO W7900 48G48GB · +6.8DGX Spark 64G64GB · +22.8Mac M5 Pro 64G64GB · +6.8
- Tight but possible
- A100 40G40GB
- Ships in
- GGUF
- A top pick for
- Best local LLM for 24GB
- Try it yourself
- Size it in the calculatorGenerate the run command
How to actually run this
Deployment guides for this model and this class of hardware.
This page's figures change when the model or the runtime does.Last updated 2026-11-11 RSS → /feed.xml