Ripple-65B Preview pricing, plans and limits
Ricursive's first public model: a dense 65B preview released as a general-purpose variant of the model behind its chip design platform.
- Best for
- Local terminal coding agents
- Not for
- Image, audio or video input
Should you use Ripple-65B Preview?
Ricursive Intelligence released Ripple-65B Preview on 4 November 2026 and reports 80.52 on Terminal-Bench 2.1 after pretraining on 12 trillion tokens (vendor-reported). It suits developers who want an open-weight coding agent they can run on their own hardware, and it competes with far larger open models such as Reflection Beam.
Best for
- Local terminal coding agents Ollama's page lists launch commands for Claude Code, OpenCode, Hermes Agent and OpenClaw, and Ricursive reports 80.52 on Terminal-Bench 2.1.
- Long codebase sessions on one machine Up to 1M tokens of documented context in a model Ricursive sizes for a single DGX Spark-class box; Ollama's readme pitches it for large codebases and long agent sessions.
Not for
- Image, audio or video input The model is text only, per Ollama's listing and Ricursive's description of the training corpus.
- Production work that cannot tolerate change Ricursive labels it a preview with relatively light post-training.
- Top agentic accuracy regardless of hardware Ricursive's own table gives Reflection Beam and Muse Spark Open higher scores on SWE-bench Verified and MCP Atlas.
Can you use Ripple-65B Preview?
- OpenRouter: preview (Pay per token through hosts Prime Intellect, Together and GMI Cloud). Model labelled Preview by Ricursive.
- Together AI: preview. Named as a host in Ricursive's launch post.
- Fireworks AI: preview. The launch post links a Fireworks model page.
- GMI Cloud: preview. The launch post links a GMI Cloud console page.
- Free plan: not published.
- Open weights: released (Apache 2.0).
Checked 11 Nov against Ricursive Intelligence's own pages.
Ripple-65B Preview pricing
- Input, per 1M tokens
- $0.10/1M
- Output, per 1M tokens
- $0.30/1M
- Cached input, per 1M tokens
- $0.03/1M
- Pricing model
- per-token
- Free tier
- No
What a real job costs
| Job | Input | Output | Total |
|---|---|---|---|
| Summarise a 20-page PDF | $0.0030 | $0.0003 | $0.0033 |
| Support reply | $0.0002 | $0.0001 | $0.0003 |
| One coding agent run | $0.020 | $0.0060 | $0.026 |
Budgets: 20-page PDF = 30k in / 1k out · Support reply = 2k in / 300 out · Coding agent run = 200k in / 20k out. Computed from the vendor's per-token prices at render time; cached-input discounts are not applied.
Prices come from OpenRouter host listings (read November 2026), not from Ricursive, which posts none: Prime Intellect at 0.10 dollars per 1M input tokens and 0.30 output (cached input 0.025), Together and GMI Cloud at 0.20 input and 0.60 output. Other hosts price differently. The open weights cost only your own compute to self-host.
Self-hosting requirements
- Self-hosted
- Yes
- Recommended GPU
- Ricursive says it fits hardware such as the NVIDIA DGX Spark; it reports 250 tokens per second on 2xB200 at batch size 8
What Ripple-65B Preview can do
Ripple-65B Preview is a 65-billion-parameter dense open-weight language model from Ricursive Intelligence, built for agentic coding and launched in preview on 4 November 2026. It supports up to 1,048,576 tokens of context, takes text input and ships under the Apache 2.0 licence.
What goes in and what comes out
Reads
- text
- tool-calls
Writes
- text
- tool-calls
- code
Capabilities
- Function callingSupported
- Tool useSupported
- Audio inputNot supported
- Audio outputNot supported
- Video inputNot supported
- VisionNot supported
- Reasoning modes
- thinking
Key features
Attention design
Ricursive says it uses 80 layers at a width of 5,120, more attention heads than usual, more sliding-window layers with a smaller window, gated attention and global NoPE layers.
Training run
The card lists 877,500 pretraining steps, 8,000 midtraining steps at 64K length and 700 reinforcement-learning steps at 128K length with a batch of 2,048.
Pretrain-only base checkpoint
The BF16 base model has no midtraining or synthetic QA data, per the card, which suits teams that want to run their own post-training.
Speculative decoding on Ollama
On Ollama's MLX engine the model uses DFlash 2 speculative decoding for faster generation, per Ollama's readme.
Known limitations
- Text only: screenshot-driven or voice agents need a second model.
- Preview with relatively light post-training, in Ricursive's words, so behaviour may change between releases.
Who it is for
- Developers who want an open-weight coding agent they can run on their own hardware
- Teams evaluating open-weight models for terminal-based coding agents
- Researchers comparing dense and mixture-of-experts coding models
Supported languages
- English
Ripple-65B Preview benchmarks
Ripple-65B Preview against its nearest rivals
SWE-bench Verified % solved, higher is better
GPQA Diamond % correct, higher is better
Standout result
Terminal-Bench 2.1 score, Ricursive-reported
Source: Ricursive's Hugging Face model card, read 11 Nov 2026Where it sits
Cheaper than 90% of the 104 GA models with a published price, and mid-pack on SWE-bench Verified (rank 20 of 40). Ranked against GA models; this record is not GA.
| Measure | This model | Rank | Median |
|---|---|---|---|
| Blended price (3:1)$ per 1M tokens | $0.15/M | #11 / 104 | peer median $1.70/M |
| SWE-bench Verified% solved | 76.5% | #20 / 40 | peer median 76.5% |
| GPQA Diamond% correct | 85.2% | #39 / 64 | peer median 87.6% |
Ranks are against GA models on Hokai that publish the same figure; ties share a rank.
A benchmark is an exam, not the job. Scores transfer unevenly between tasks, so weigh the one closest to your workload and read every figure with its source.
Strengths and limits
Strengths
- Permissive licence: commercial use is allowed, per Ollama's readme and the Hugging Face card, so a team can self-host and fine-tune it.
- Ships with the pieces to reproduce or tune it: a base checkpoint, an NVFP4 checkpoint and a speculative-decoding checkpoint sit beside the instruct model on Hugging Face.
- Early interest from outside Ricursive: 487 likes on its Hugging Face page and 1,717 Ollama downloads, both read on 11 November 2026, a week after release.
- Small enough to run locally: Ricursive says it fits hardware such as the NVIDIA DGX Spark, and ships an NVFP4 checkpoint.
Limits
- Every score on this page is Ricursive-reported. Hokai did not run the model, which needs tens of gigabytes of memory, and read no independent leaderboard entry for it.
- Text only, so teams that need image, audio or video input need a second model.
- On OpenRouter, read 11 November 2026, the host with the lowest posted rate showed 5.72 s latency and 23 tokens per second, against 0.61 s and 80 tokens per second on Together.
- A preview with light post-training, so behaviour may change before a stable release.
Ripple-65B Preview alternatives
| Model | Lab | Context | In / out per 1M | Reads and writes |
|---|---|---|---|---|
| Ripple-65B Preview This page | Ricursive Intelligence | 1M | $0.10 / $0.30 | text → text, code |
 BeamPick Ripple if you want a 65B dense model for one machine; pick Beam if you have server GPUs and want its higher SWE-bench Verified score, 81.6 against 76.5 on Ricursive's card. | Reflection AI | 1M | $1 / $3 | text → text |
 Muse Spark OpenPick Ripple for a smaller dense model you can run locally; pick Muse Spark Open if tool-use accuracy matters more, with 89.2 on MCP Atlas against Ripple's 71.67 on Ricursive's card. | Meta AI | 1M | -- | text, image, video, audio → text |
 Claude Fable 5Superseded by Fable 5.1 and Fable 5.2 but still sold on the Claude API: Anthropic's first generally available Mythos-class model, released June 2026, with a 1M-token context. | Anthropic | 1M | $10 / $50 | text, image, pdf → text |
Ripple-65B Preview specs
- Context window
- 1.0M tokens
- Max output
- 131K tokens
- Mixture of experts
- No
- Parameters
- 65B (all active per token)
- Architecture
- Transformer (dense), 80 layers, sliding-window plus global attention
- VRAM requirements
- Not published; Ollama's builds are 39 GB to 46 GB on disk
Training data
Pretrained from scratch on a 12T-token, text-only, predominantly English and code corpus at 8K sequence length, then midtraining, long context, SFT, long-context SFT and mixed-domain RL, per Ricursive's Hugging Face card.
Weight formats
- safetensors
- BF16
- NVFP4
API endpoints
| Name | Endpoint | Auth |
|---|---|---|
| OpenRouter | https://openrouter.ai/ricursiveai/ripple-65b-preview | API key |
| Together AI | https://www.together.ai/models/ripple-65b-preview | API key |
| Hugging Face weights | https://huggingface.co/ricursiveAI/Ripple-65B-Preview | None for download |
| Ollama library | https://ollama.com/library/ripple-preview | None for download |
SDKs
- Python
- JavaScript
Known quirks
Show 3 known quirks
Context size differs by page
Ollama's tag table shows 2M for ripple-preview, while Ollama's readme, Ricursive and OpenRouter say up to 1M.
Workaround: Plan around 1,048,576 tokens, set the context length explicitly in your runner, and test any longer prompt before relying on it.
Agentic scores depend on the test setup
Ricursive's card says Terminal-Bench 2.1 scores were generally lower under the official Terminus 2 setup at 128K sequence length with summarisation.
Workaround: Re-run your own task set on the agent setup you will ship before trusting any published score.
Light post-training in a preview
Ricursive says the preview had relatively light post-training with a focus on agentic coding, so non-coding behaviour is less tuned.
Workaround: Pin the exact checkpoint you test and re-evaluate when Ricursive releases an update.
Ripple-65B Preview facts
- Licence
- Apache 2.0 · Open source
- Released
- 04 Nov 2026
- Current version
- Ripple-65B-Preview
- Family
- Ripple
- Variant
- Chat
- Deployment
- API · Self-hosted · Hugging Face · Ollama · OpenRouter · Together AI · Gmi Cloud · Fireworks AI
Safety controls
- System prompt
- Yes
- Constitutional AI
- No
- Content filter API
- No
- Moderation endpoint
- No
Version history
Ripple-65B-Preview · 04 Nov 2026
Initial preview: instruct model in BF16 and NVFP4, a BF16 base model and a DFlash2 checkpoint.
What changed
- Release Ripple-65B Preview released as open weights source
Questions about Ripple-65B Preview
What does Ripple-65B Preview cost per million tokens?
Ricursive publishes no rate card of its own. OpenRouter lists three hosts, read on 11 November 2026: Prime Intellect at $0.10 input and $0.30 output per 1M tokens, Together at $0.20 and $0.60, and GMI Cloud at $0.20 and $0.60. Cached input is $0.025, $0.02 and $0.05 per 1M on those three hosts in that order. Self-hosting the Apache 2.0 weights costs only your own hardware.
Is Ripple-65B Preview open source, and can I run it for free?
The weights are released under Apache 2.0, which Ollama's readme says allows commercial use. Ricursive calls it an open-weight model and did not publish its pretraining corpus on the pages read. You can run it locally at no licence cost through Ollama as ripple-preview, or download the BF16 and NVFP4 checkpoints from Hugging Face. No free hosted tier was listed on the pages read.
Does Ripple-65B Preview support 1M or 2M tokens of context?
The documented limit is 1,048,576 tokens, per the Hugging Face card, and OpenRouter lists 1.0M. Ollama's tag table shows 2M for the same model while its own readme says up to 1M, and Ricursive's training table includes long-context stages at 2M sequence length. Treat 1M as the supported figure and test anything longer yourself.
How does Ripple-65B Preview compare with Reflection Beam for coding agents?
On Ricursive's card, Ripple scores 80.52 on Terminal-Bench 2.1 against 80.1 for Beam as Reflection reports it, while Beam leads on SWE-bench Verified, 81.6 to 76.5, and on SWE-bench Pro v1.0, 66.1 to 61.42. Ripple has 65B dense parameters; Beam has 501B in total and 23B active. Both sets of numbers are vendor claims, and Hokai read no independent board for either.
What hardware does Ripple-65B Preview need to run locally?
Ricursive says it fits consumer hardware such as the NVIDIA DGX Spark and provides an NVFP4 checkpoint for local use. Ollama's builds are 39 GB to 46 GB on disk, which sets the memory floor before the context cache. Ricursive reports 250 tokens per second on two B200 GPUs at batch size 8, before quantisation or custom kernels.
About Ripple-65B Preview
Ripple-65B Preview is the first model from Ricursive Intelligence, a dense language model with 65 billion parameters, put out as a preview on 4 November 2026. Ricursive describes it as a general-purpose variant of the proprietary model behind its AI chip design platform, pretrained from scratch on 12 trillion tokens of text and code. Every parameter is active on every token, unlike the mixture-of-experts designs of Reflection Beam and Muse Spark Open, the two open-weight models Ricursive benchmarks it against. The release covers an instruction-tuned model in BF16 and NVFP4, a BF16 base model and a DFlash2 checkpoint used for speculative decoding.
Read the full background
The scores below are all Ricursive-reported, taken from its Hugging Face model card (dated in the benchmark list). It reports 80.52 on Terminal-Bench 2.1, 76.5 on SWE-bench Verified, 61.42 on SWE-bench Pro v1.0, 85.2 on GPQA Diamond, 84.2 on MMLU-Pro and 79.2 on IFBench. On the same card, the Terminal-Bench 2.1 scores printed for the other labs' models are those labs' own: 80.1 for Beam and 76.4 for Muse Spark Open, so the 0.4-point gap to Beam is a self-reported comparison, not an independent result. Ricursive also warns that agentic scores shifted with the test setup: most models scored lower under the official Terminus 2 setup. Hokai found no independent leaderboard entry for the model on the day of writing.
The context window is 1,048,576 tokens (1Mi) on the Hugging Face card and 1.0M on OpenRouter. Ollama's own page disagrees with itself (2M in its tag table, up to 1M in its readme), so only 1M counts as the documented limit. Ricursive lists no maximum output length, while OpenRouter's API gives 131,072 tokens for a single completion.
The model takes text in and gives text out. Ollama lists the input as text, and Ricursive describes the pretraining corpus as text-only, mostly English and code. OpenRouter lists native tool calling, and Ollama tags the model for tools and thinking. Ricursive documents no image, audio or video input.
The weights are on Hugging Face under Apache 2.0, and Ollama's readme says commercial use is allowed. Ollama serves it as ripple-preview in builds of 39 GB to 46 GB and lists launch commands for Claude Code, OpenCode, Hermes Agent and OpenClaw. Ricursive's launch post names four hosted routes: Together AI, GMI Cloud, Fireworks and OpenRouter. See OpenRouter, Together AI and Fireworks AI for those hosts.
On safety, the card reports StrongREJECT 97.5, FORTRESS Adversarial 83.8 and FORTRESS Benign 95.5, all run in house. Hokai read no system card, red-team partner list or usage policy for it, and Ricursive says post-training was relatively light, with a focus on agentic coding.
Pick something else if accuracy matters more than footprint. Ricursive's own table gives Reflection Beam 81.6 on SWE-bench Verified against Ripple's 76.5, and Muse Spark Open 89.2 on MCP Atlas against Ripple's 71.67. Both are far larger mixture-of-experts models, at 501B and 816B total parameters on Ricursive's card. Anything with images, audio or video needs a different model, and a preview with light post-training is a weak base for production work that cannot tolerate behaviour changes.
Ricursive says it pretrained at 8K sequence length, midtrained on 200 billion tokens, then ran long-context, supervised fine-tuning and mixed-domain reinforcement-learning stages. It says it bought no post-training data and built its own synthetic data pipelines, and it puts its full research and development cost at 32 million dollars, a figure it estimates itself. The training cutoff is not published.
Sources and guides
Official sources
- Ripple-65B Preview model page Vendor
- Licence text Licence
Specifications on this page (context window, token prices, benchmark scores, licence) are read from Ricursive Intelligence's own model card and documentation. The verdict is Hokai's reading of them, not the provider's claim.
Lists that include Ripple-65B Preview
The provider
More AI Models on Hokai
- Claude Fable 5.1Superseded by Claude Fable 5.2 in November 2026 but still sold on the Claude API: Anthropic's second Mythos-class model, released Sept 1, 2026, for long-horizon coding and research.
Gemini Decision 1Google DeepMind's decision model scores many questions over one multimodal input of up to 262,144 tokens and bills zero output tokens.
Shieldstral 1.0Mistral's open-weight multimodal safety classifier, released 4 August 2026, that screens text and images against policies you write.
- Claude Mythos 5.2Anthropic's November 2026 Mythos model: Fable 5.2's weights with relaxed cyber and biology safeguards for verified teams, with a 1M-token context.
ERNIE-5.5-Flash-PreviewBaidu's text-only preview mixture-of-experts model, served through Qianfan on OpenRouter with a 1M-token context window and tool calling.
Gemini 4 CarbonGoogle DeepMind's Gemini 4 Carbon, open to paying Gemini API developers since 10 November: a reasoning model for agentic coding, research and cyber defense.