← All guides
Models November 5, 2026

Can I Run Ripple-65B Locally? (2026): Yes, 38.8 GB Q4

Ricursive Intelligence released Ripple-65B Preview on 4 November 2026 under Apache 2.0. It is a dense 65B, not a mixture of experts, so every token reads all the weights. The Q4_K_M GGUF is 38.8 GB. That rules out one 24GB card and one 32GB card at Q4, and it fits two RTX 3090s, a 48GB workstation card, a 64GB Mac, a Strix Halo box or a DGX Spark. Here are the byte-exact sizes, the llama.cpp build that loads it, and every speed number published so far.

Bottom Line

  • Can I run Ripple-65B locally? Yes, from 48GB up. The community Q4_K_M GGUF is 38.8 GB. Two RTX 3090s, a 48GB workstation card, a 64GB Mac, a 128GB Strix Halo box or a DGX Spark hold it. One 24GB card does not.
  • It is dense, not MoE. All 65B parameters read on every token. On a 256 to 273 GB/s box that means single-digit tokens per second without a draft model. On a 1.2 TB/s Mac Studio it means about 22 tok/s (community-reported).
  • You need a llama.cpp build from 5 November 2026 or later. PR 31692 added the ricursive architecture and merged that day. Ollama and LM Studio did not load the files at launch.
  • Sizes to plan around: Q3_K_M 28.9 GiB, Q4_K_M 38.8 GB, Q6_K 49.5 GiB, Q8_0 68.8 GB, NVFP4 57.2 GB, BF16 130 GB.
  • Best buy if you want this model: two used RTX 3090s at Q4_K_M. Cheapest working path: a 64GB Mac Studio M5 Max. Slowest working path: Strix Halo or one DGX Spark without speculative decoding.

Ready to buy? See the tested hardware list with current prices.

What Ripple-65B Preview Is

Read from the model card and config.json on 5 November 2026.

Ripple-65B Preview
MakerRicursive Intelligence (US)
Released4 November 2026
Parameters65B, dense
Layers80: 64 sliding-window (1,024 tokens) + 16 full-attention, one full layer in every five
Attention80 heads, 8 KV heads, head dim 128, gated attention
Hidden size5,120 (narrow for the size; intermediate 40,960)
Vocabulary100,352
Training12T tokens from scratch, 8K sequence, BF16 throughout
Context1M claimed on the card; max_position_embeddings is 2,097,152
CheckpointsBF16 instruct (130 GB), NVFP4 experimental (57.2 GB), BF16 base, DFlash2 draft model (4.04B)
LicenseApache 2.0
Self-reported scoresTerminal-Bench 2.1 80.5, SWE-bench Verified 76.5, SWE-bench Pro v1.0 61.4, GPQA Diamond 85.2

Why it trends. Ricursive calls it the state of the art for US open-weight models on agentic coding, above Muse Spark Open 816B and Beam 501B on Terminal-Bench 2.1, and says the full R&D cost was $32M. Those numbers are the company’s own. What matters for this page is the shape: a dense 65B that fits a desk, with a context window the card puts at 1M.

The NVFP4 build is not a full 4-bit model. The card says it quantizes the MLP layers to NVFP4 and keeps attention in BF16. That is why it is 57.2 GB and not about 35 GB. It targets Blackwell GPUs and the DGX Spark, not a 3090.

The Arithmetic: What Fits Where

File sizes below are read from the Hugging Face file trees on 5 November 2026. Hugging Face shows GB; some README tables show GiB. Both are given where both exist.

QuantFileSizeSource
IQ3_XXSprometheusAIR25.38 GiB (27.3 GB)community
Q3_K_Mmrfakename28.9 GiB (31.0 GB)community
Q4_K_Mmrfakename, Chungulus38.8 GB (36.1 GiB)community
4.64bpwddh037.6 GBcommunity
MLX 4-bitfrogggiasabout 37 GBcommunity
Q5_K_Mmrfakename42.6 GiBcommunity
Q6_Kmrfakename49.5 GiBcommunity
Q6_K_XLprometheusAIR52.76 GiBcommunity
NVFP4ricursiveAI57.2 GBofficial, experimental
Q8_0mrfakename, Chungulus, ddh068.8 GB (64.1 GiB)community
BF16ricursiveAI130 GBofficial

Add the KV cache. This model is cheaper on context than a plain dense 65B. Only 16 of 80 layers keep a full-length cache. The other 64 keep a fixed 1,024-token window. Derived from config.json (8 KV heads × 128 dim × 2 for K and V × 2 bytes): about 64 KB per token at f16 for the full layers, so about 8 GB at 128K context, plus a fixed 256 MB for the sliding layers. With -ctk q8_0 -ctv q8_0 halve that. The numbers are derived, not measured; check llama-server output on your machine.

Your machineMemoryLargest sensible quantFits?
RTX 3090 or 409024GBnone; IQ3_XXS is 27.3 GBNo
RTX 509032GBIQ3_XXS (27.3 GB), about 4 GB left for KVBarely, short context only
2x RTX 309048GBQ4_K_M (38.8 GB)Yes, about 9 GB for KV
RTX A6000 or RTX PRO 500048GBQ4_K_MYes
Mac Studio M5 Max48GBQ3_K_M (31.0 GB) onlyTight
Mac Studio M5 Max64GBQ4_K_M or MLX 4-bitYes
RTX PRO 6000 Blackwell96GBQ8_0 (68.8 GB)Yes
Strix Halo mini PC128GBQ8_0Yes, slow
NVIDIA DGX Spark128GBQ8_0 or NVFP4Yes, slow without a draft model
Mac Studio M5 Ultra256GBQ8_0 or MLX 8-bitYes, the fast unified-memory path

Speed: Every Number Published So Far

None of these are our measurements. Each is community-reported by the person who published the build, compiled in a 5 November 2026 write-up by Alan West on DEV, who states he did not run the model. Treat them as a first read, not a benchmark.

HardwareBuildReported speed
DGX Spark (1x GB10, 128GB)NVFP4 + DFlash2 draft, community Spark server kit20-36 tok/s on code (average 26.5), about 9 on prose; plain decode estimated at about 5 tok/s
Mac Studio M5 Ultra 256GBMLX 4-bit (37 GB)22.0-23.3 tok/s decode; 535-565 tok/s prefill
RTX PRO 6000 96GBQ6_K_XL GGUF, short prompt26.2 tok/s
RTX PRO 6000 96GBIQ3_XXS GGUF, short prompt46.0 tok/s
RTX PRO 6000 96GBQ6_K_XL, 32K prompt, q8_0 KV1,448 tok/s prefill, 23.9 tok/s decode
RTX PRO 6000 96GBQ6_K_XL, 1.03M prompt, q8_0 KV226 tok/s prefill (76 minutes), 6.7 tok/s decode
GH200 96GB (datacenter)Q4_K_M GGUF40.6 tok/s
2x B200 (datacenter)BF16, batch 8, SGLang250 tok/s total, per Ricursive

Read the Spark line carefully. The 26.5 tok/s is with speculative decoding through the 4B DFlash2 draft model, which drafts 16 tokens at a time. The kit author’s own estimate for plain decoding is about 5 tok/s. That matches the arithmetic: a 273 GB/s box reading 57 to 69 GB of weights per token lands between 4 and 5 tok/s. The draft model is doing the work.

Derived ceilings for the hardware with no published number. Decode speed on a dense model is bounded by memory bandwidth divided by the weight bytes read per token. These are upper bounds. Real numbers land at 60 to 75 percent of them.

HardwareBandwidthQ4_K_M ceiling (38.8 GB)
2x RTX 3090 (layers split, one card active at a time)936 GB/sabout 24 tok/s
RTX A6000 48GB768 GB/sabout 20 tok/s
RTX PRO 5000 48GB1,344 GB/sabout 35 tok/s
Mac Studio M5 Max 64GB614 GB/sabout 16 tok/s
Mac Studio M5 Ultra1.2 TB/sabout 31 tok/s (measured 22-23 on MLX 4-bit, 70-75% of ceiling)
Strix Halo 128GB256 GB/sabout 6.6 tok/s
DGX Spark273 GB/sabout 7 tok/s

The M5 Ultra row is the sanity check: the one measured unified-memory number sits where the ceiling math says it should.

How to Run It

llama.cpp. Mainline support arrived in PR 31692, merged 5 November 2026. A release build from 5 November or later loads the GGUFs above. Before the merge, every GGUF README said the same thing: stock llama.cpp, Ollama and LM Studio do not load this architecture. Those downstream tools pick up the change only when they bump their llama.cpp. If llama-server reports an unknown architecture ricursive, your build is older than the merge.

# two 3090s, Q4_K_M, 64K context, q8_0 KV cache
llama-server -m Ripple-65B-Preview-Q4_K_M.gguf \
  -ngl 99 -c 65536 -ctk q8_0 -ctv q8_0 -fa on --port 8080

The split GGUFs (Q8_0 ships as two files) load from the first shard. Point -m at -00001-of-00002.gguf and llama.cpp finds the rest.

MLX on a Mac. Use the community 4-bit or 8-bit conversion with mlx-lm 0.32.2 or newer and --trust-remote-code. That flag runs a custom ricursive.py of about 280 lines from the repo. Read it before you run it.

DGX Spark. Ricursive’s own path is the NVFP4 checkpoint served with SGLang. The DFlash2 draft model needs a runtime that supports both Ripple and DFlash2DraftModel; the card’s example is --speculative-algorithm DFLASH. A community Spark server kit wraps this. Without the draft model, expect about 5 tok/s.

vLLM or SGLang on a 48GB card. The BF16 checkpoint is 130 GB and does not fit. The NVFP4 build is 57.2 GB and does not fit either. On 48GB the GGUF is the only route today.

What to Buy for Ripple-65B

Prices are US street as of November 2026, from this site’s price reference. The used 3090 range is the resaleprices.com read of 5 November 2026. Everything else is dated in the Sources list.

RouteMemoryWhat it runsPrice as of Nov 2026Link
Two used RTX 309048GBQ4_K_M, about 24 tok/s ceiling$1,500-1,650 each, $3,000-3,300 the pairRTX 3090 24GB
Used RTX A600048GBQ4_K_M in one slot, about 20 tok/s ceiling$2,600-3,800 used (Sep 2026 read)RTX A6000 48GB
Mac Studio M5 Max 64GB64GBQ4_K_M or MLX 4-bit, about 16 tok/s ceiling$3,499 (apple.com, Oct 2026)no link; order from Apple
Strix Halo mini PC128GBQ8_0, about 6.6 tok/s ceiling$3,499.99 GMKtec EVO-X2 (Sep 2026 read)Ryzen AI Max+ 395 128GB box
NVIDIA DGX Spark128GBNVFP4 + DFlash2, 20-36 tok/s community-reported$6,950 (Oct 2026)DGX Spark 128GB

Our pick: two used RTX 3090s. For this one model it is the cheapest 48GB that is also fast. A pair costs about what one 64GB Mac Studio costs and its ceiling is 50 percent higher. The costs are a 1,200W PSU, 700W at the wall, and the used-market risk covered in How to Buy a Used RTX 3090.

Buy the Mac if the machine is also your laptop replacement or has to be silent. The M5 Max 64GB runs Q4_K_M at about 16 tok/s on paper. The 256GB M5 Ultra is the only unified-memory box with a published number above 20 tok/s, at $9,499.

Do not buy a Strix Halo box for this model. At 256 GB/s a dense 65B is a 6 tok/s experience. Those boxes earn their price on MoE models like gpt-oss 120B, where only a few billion parameters read per token. See Best Models for Ryzen AI Max+ 395.

The DGX Spark is a special case. Same bandwidth problem as Strix Halo, but Ricursive aimed the NVFP4 build and the DFlash2 draft model at it. The 26.5 tok/s community number depends on speculative decoding and on code prompts, where draft acceptance is high. Prose drops to about 9.

The Honest Caveats

  1. Every speed number above is secondhand. The compiler of the table says he did not run the model. We have not either. Expect the first independent benchmarks within a week.
  2. “Preview” means the card can change. The NVFP4 build is labelled experimental. The BF16 README was edited within the hour before we read it.
  3. Dense 65B against Qwen3.8 27B. The 27B fits one 24GB card and runs 2 to 3 times faster on the same bandwidth. Ripple earns its cost only if the agentic-coding scores hold up in your own harness. Run both on the same task before you buy a second GPU.
  4. The 1M context is a card claim. The one long-context receipt is the RTX PRO 6000 table: 1.03M tokens of prefill took 76 minutes at 226 tok/s. That is a batch job, not a chat.
  5. Ollama and LM Studio lag. The GGUF READMEs say they do not load the files. Until each tool ships a llama.cpp from 5 November or later, use llama-server directly.

Sources

See Also

Before you order parts, check the tested hardware list for current prices by tier.

Read next

Can I Run Underdog Saluki 27B on 8GB VRAM? (2026): Not Fully
Underdog Saluki 27B is a 7.36 GiB GGUF of Qwen3.8-27B. On an 8GB GPU the weights fit but context does not. Byte-exact sizes, KV math, the -ngl values that work, and what 12GB, 16GB and Mac buyers get.
Bonsai 2 27B on RTX 3060 12GB: Fits, Needs a Fork (2026)
Ternary Bonsai 2 27B is a 1.72-bit Qwen3.8-27B that fits a 12GB RTX 3060, and even an 8GB card. It needs the PrismML llama.cpp fork, not Ollama or LM Studio. File sizes, KV cache math for 12GB, vendor-reported quality, and how it compares to Qwen 3.8 27B on a 3090.
DeepSeek V4.1 Flash GSQ-RCO Quants (2026): 12GB GPU, No Mac
ISTA-DASLab shrank DeepSeek V4.1 Flash 1017 to 1.75-2.25 bits per weight. The backbone is 112-145 GiB, the Engram table is 103 GiB on SSD, and the files run today only in Strata on a 12GB+ NVIDIA card. Stock llama.cpp cannot load them, and no Mac runs them yet.
gpt-oss 120B vs 20B (2026): Which One Should You Run?
gpt-oss 20B fits a 16GB card at short context and does not fit one at its full 128K window — the KV cache is 3.0 GiB and the weights leave about that much room. The 120B needs 96GB or a 128GB unified box. Both figures come from the models' own config.json.