# Ripple-65B Preview (slim)

> Ricursive Intelligence's 65B dense model, pretrained from scratch and post-trained for agentic coding, on Hugging Face in BF16 and NVFP4 under Apache 2.0. Its custom architecture needs Ricursive's own model code to run.

- Full: https://www.anchorterminal.com/tools/ripple-65b-preview.md (~9,000 tokens) · this version ~1,880 tokens · JSON https://www.anchorterminal.com/tools/ripple-65b-preview.json · canonical https://www.anchorterminal.com/tools/ripple-65b-preview
- Index: https://www.anchorterminal.com/llms.txt · API: https://www.anchorterminal.com/api/v1/index.json · MCP: https://www.anchorterminal.com/mcp · Updated: 2026-11-04 · pricing https://www.anchorterminal.com/tools/ripple-65b-preview#pricing · reviews https://www.anchorterminal.com/tools/ripple-65b-preview#reviews
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)

**D · 53.3/100 · rank #103 of 121 open-weight models · #78 in Open-weight models · not agent-ready · confidence low**


Assessment: A dense 65B coding model from a new lab, pretrained from scratch and released today under Apache 2.0, that Ricursive says leads U.S. open-weight models on Terminal-Bench 2.1. Nobody has checked that yet, and the custom architecture runs only through Ricursive's own SGLang plugin on Blackwell GPUs, with a 57 GB NVFP4 build as the smallest.

## Facts

- Kind: Open-weight model · vendor: Ricursive Intelligence · category: Open-weight models · legal entity: Ricursive Intelligence, Inc. · provenance 36/100
- Packages: huggingface `ricursiveAI/Ripple-65B-Preview`, huggingface `ricursiveAI/Ripple-65B-Preview-NVFP4`
- Auth: None · pricing: Free · x402: no · licence: Apache-2.0
- Probe metrics: not measured yet (probes haven't run)
- Surface graded: The open weights at huggingface.co/ricursiveAI/Ripple-65B-Preview and its NVFP4 build, read from the model cards, the Hub's model info and file pages, config.json, tokenizer_config.json, the licence, the server README and the Spark loader on 4 November 2026. Nothing was downloaded or run. Hosted copies aren't graded
- Model: RicursiveForCausalLM, 64,782,689,360 BF16 parameters per the Hub. Dense, 80 layers, hidden size 5,120, 80 attention heads and 8 KV heads, sliding window 1,024 on four layers of every five, vocabulary 100,352. Text in, text out. Pretrained from scratch, not built on another model, starting from Olmo 3's model ladder
- Builds: BF16 instruct (16 safetensors shards, 129.6 GB), NVFP4 instruct (6 shards, 57.2 GB, NVFP4 MLPs and BF16 attention, which its card calls experimental and conservative), a BF16 pretrain-only base model and a DFlash2 draft model for speculative decoding, each in its own repository. No GGUF or other build from Ricursive
- Files: Safetensors with a SHA-256 for every LFS file on the Hub, chat_template.jinja, tokenizer files, configuration_ricursive.py and modeling_ricursive.py (loaded through auto_map), and inference folders of Python (serve.py, model.py, ripple_format.py, and spark.py with a notebook in the NVFP4 repository). No pickle files
- Runtimes: Ricursive's server registers its own model.py and parsers into SGLang 0.5.19 on two B200 GPUs, and its Spark loader does the same for the NVFP4 build. Transformers loads it only through trust_remote_code. No released llama.cpp, Ollama, vLLM, SGLang, LM Studio or Transformers version was found to run the ricursive architecture with its own code
- Context: Up to 1M tokens per the card, after long-context stages at 2M-token sequences. config.json allows 2,097,152 positions, tokenizer_config.json says 131,072, and the server defaults to 16K
- Tools and reasoning: The tools-aware ChatML template takes tools in tools tags and renders one or more tool_call blocks per turn. With Ricursive's ripple parsers the server returns reasoning_content, content and tool_calls separately. The template has no thinking switch and no reasoning-effort levels
- Hardware: The server README targets two B200 GPUs at TP2 with a 16K context. The cards say the model fits an NVIDIA DGX Spark with the NVFP4 build, and give 250 tokens a second on 2 × B200 at batch 8 in BF16
- Training data: 12T tokens of text-only, predominantly English and code, then 200B midtraining, 6.5B long-context, 183B SFT and 13.5B long-context SFT tokens and RL, per the card's stage table. No sources, filtering or knowledge cut-off
- Ricursive's figures: Vendor-reported at full BF16 and maximum effort. Terminal-Bench 2.1 80.5, SWE-bench Verified 76.5, SWE-bench Pro v1.0 61.42, DeepSWE 1.1 44.3, MCP Atlas 71.67, GPQA Diamond 85.2, MMLU-Pro 84.2, IFBench 79.2, AA-LCR 75.0. Safety FORTRESS adversarial 83.8, FORTRESS benign 95.5, StrongREJECT 97.5. Harnesses are OpenCode or Terminus 2

- Scores: Capability pending, Reach 45, Risk 55, Agent interface 75, Documentation and upkeep 49 · total over the 7 assessed categories
- Why:  · Reach, Memory 10 of 30, the smallest 4-bit build is Ricursive's NVFP4 at 57.2 GB, 68.7 GB with a fifth added, which is over 64 GB and fits our refe… · Risk, File safety 15, safetensors only, less 10 because the documented way to run it loads Ricursive's Python. · Agent interface, Chat template 20, chat_template.jinja in the repository. · Documentation and upkeep, Model card 14 of 20, no limitations and no knowledge cut-off.
- Sources: 20, open questions: 6, both in the full twin
- Capabilities: model.open-weights
- JSON: https://www.anchorterminal.com/api/v1/tools/ripple-65b-preview.json
- Verify (for the vendor): the badge `https://www.anchorterminal.com/badges/ripple-65b-preview.svg` or a link to https://www.anchorterminal.com/tools/ripple-65b-preview from a page on ricursive.com or one of its subdomains, then `POST https://www.anchorterminal.com/api/v1/verify` `{"slug", "url"}` or `verify_listing` at /mcp; re-checked weekly, no effect on the grade. Snippets in the full twin.
- Support: https://huggingface.co/ricursiveAI/Ripple-65B-Preview/discussions; more in the full twin

## Before you call it

1. Run Ricursive's launcher from inference/server with `uv run --script serve.py` on two B200 GPUs; it listens on 127.0.0.1:30000 with a 16K context by default
2. The launcher turns on `--reasoning-parser ripple` and `--tool-call-parser ripple`; without them tool calls come back as raw text inside tool_call tags
3. For `tool_choice: "required"` set `strict: true` and `additionalProperties: false` on each function, which the README says stopped extra argument fields
4. Leave room in `max_tokens` for reasoning as well as the answer, since thinking is always on
5. Use chat_template.jinja, not the template inside tokenizer_config.json, which has no tool support

## Connect

```bash
hf download ricursiveAI/Ripple-65B-Preview --local-dir ./ripple-65b
hf download ricursiveAI/Ripple-65B-Preview-DFlash2 --local-dir ./ripple-dflash2
cd ripple-65b/inference/server
CUDA_VISIBLE_DEVICES=0,1 uv run --script serve.py \
  --model-path ../.. \
  --speculative-draft-model-path ../../../ripple-dflash2
```

```bash
curl http://127.0.0.1:30000/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{"model": "ripple", "messages": [{"role": "user", "content": "Hello"}], "max_tokens": 128}'
```

## Similar tools

| Tool | Grade | Score | Shared capabilities | Slim |
| --- | --- | --- | --- | --- |
| Qwen3-Embedding-8B | BB | 75.4 | model.open-weights | https://www.anchorterminal.com/tools/qwen3-embedding-8b.min.md |
| Gemma 4 31B IT | BB | 74.1 | model.open-weights | https://www.anchorterminal.com/tools/gemma-4-31b-it.min.md |
| Qwen3-Embedding-4B | BB | 73.6 | model.open-weights | https://www.anchorterminal.com/tools/qwen3-embedding-4b.min.md |
| Gemma 4 12B IT | BB | 72.3 | model.open-weights | https://www.anchorterminal.com/tools/gemma-4-12b-it.min.md |
| Gemma 4 26B A4B IT | BB | 72.2 | model.open-weights | https://www.anchorterminal.com/tools/gemma-4-26b-a4b-it.min.md |


## Panel reviews (0, desk reviews from public material, no calls made)