What is Ripple-65B Preview?
Ripple-65B Preview is the first public model from Ricursive Intelligence, the Palo Alto chip-design lab founded by Anna Goldie and Azalia Mirhoseini, the former Google researchers behind AlphaChip. It is not related to Recursive, Richard Socher's London lab that makes Flint. Released on 4 November 2026 under Apache 2.0, it is a dense Transformer with 64.8 billion parameters, pretrained from scratch on 12 trillion tokens of mostly English text and code, with a stated context of up to 1M tokens. Ricursive calls it a general-purpose variant of the proprietary model behind its chip-design platform.
The architecture is thin and deep: 80 layers on a 5,120-wide model, 80 attention heads, and four sliding-window layers (1,024-token window) for every full-attention layer. Post-training was deliberately light and aimed at agentic coding. Ricursive says the whole project, from scaling experiments to the final run, cost $32 million, and that a team of one to five people trained it.
Four checkpoints are on Hugging Face: the BF16 instruct model, an experimental NVFP4 build, the BF16 base model and a DFlash2 draft model for speculative decoding. On OpenRouter, GMICloud serves it in BF16 at $0.20 per million input tokens and $0.60 per million output tokens.
Is Ripple-65B Preview worth it?
Ripple-65B Preview's claim is unusual for its size: on Ricursive's own table it scores 80.5 on Terminal-Bench 2.1, ahead of Reflection Beam (501B, 80.1), Muse Spark Open (816B, 76.4), Inkling (975B, 63.8) and Nemotron 3 Ultra (550B, 56.4). Rerun on the official Terminus 2 harness, Ripple holds 80.5 while Beam falls to 70.8, and on SWE-bench Pro v2.0 Hard it posts 79.1 against Beam's 76.5. No independent lab had reproduced any of this by 5 November 2026; Artificial Analysis had no entry.
Read the whole table and the lead narrows. Ripple tops three of the 22 instruct rows, all of them agentic coding. It trails Beam on SWE-bench Verified (76.5 against 81.6), SWE-bench Pro v1.0 (61.42 against 66.1) and MCP Atlas (71.67 against 80.8), and Muse Spark Open on DeepSWE 1.1 (44.3 against 60.0). Many rival scores are Ricursive's in-house runs, marked in italics, including Inkling's 2.0 on SWE-bench Pro v2.0 Hard, a figure that looks more like a broken harness than a weak model. The card itself calls the comparison best-effort and says agentic scores are hard to reproduce.
The base model shows where the light post-training starts from. Ricursive's own runs put it at 83.7 on MMLU and 70.9 on HumanEval, below Nemotron 3 Ultra's 89.1 and 83.8, and at 10.1 on SimpleQA-Verified, so it carries little factual recall. Ricursive puts Nemotron's lead down to synthetic data, but the practical reading is that Ripple's strength sits in coding loops, not general knowledge.
Dense is the real trade-off. Every token passes through all 64.8 billion parameters, where Beam activates 23 billion, so the BF16 weights fill about 130 GB and Ricursive's server recipe needs two B200s, quoting 250 tokens a second at a batch of 8. The experimental NVFP4 build is 57.2 GB and targets a 128 GB DGX Spark, but its script fixes the context at 131,072 tokens, an eighth of the advertised 1M. For a team that can afford two data-centre GPUs and wants Apache 2.0 terms, Ripple is the smallest model in this score range; on a single workstation card it does not fit at all.
Where it earns its place
A platform engineer at a 300-person software company with a pair of B200s
Serve a coding-agent model in-house so source code never leaves the network
Run Ricursive's serve.py with SGLang 0.5.19 on two B200s and point the agent harness at the local endpoint; the README suggests 8 concurrent requests per replica, and a second pair of GPUs adds a second replica. Model licence costs nothing under Apache 2.0.
An engineering manager comparing hosted models for a 25-developer team's coding agent
Cut per-token spend on long terminal sessions
At $0.20 per million input and $0.60 per million output on OpenRouter, 100 million input and 10 million output tokens a month comes to about $26. Run it on the team's own tickets first: the headline scores are self-reported and the only listed host is GMICloud.
An ML researcher at a university lab studying post-training
Start reinforcement-learning experiments from a strong base model with a permissive licence
The BF16 base checkpoint is pretrain-only, with no midtraining or synthetic question-answer data, and ships under Apache 2.0, so results can be published and the weights redistributed. Ricursive documents each training stage, from 12 trillion pretraining tokens to 700 RL steps.
Before you commit
- Setup
- For servers, download the BF16 weights and the DFlash2 draft, then run serve.py with uv on two B200s under Linux and CUDA 13; it listens on port 30000 with a 1,048,576-token context. Loading with Transformers needs trust_remote_code because the model uses its own modeling_ricursive.py.
- Integrations
- OpenAI-style chat with reasoning returned in reasoning_content and structured tool_calls through Ricursive's ripple parsers in SGLang. OpenRouter accepts tools, tool_choice and response_format. There is no thinking on or off switch and no reasoning-effort setting.
- Data and privacy
- Apache 2.0 for all four checkpoints, with no revenue cap or use restrictions beyond the licence. Self-hosted prompts stay on your hardware; hosted terms are GMICloud's, and OpenRouter offers no EU in-region routing for it.
- Cost at scale
- BF16 weights are about 130 GB by our arithmetic from the 64.8 billion parameters, so plan on two B200s; OrcaRouter estimates that pair at about $4,800 a month. The NVFP4 files total 57.2 GB for a 128 GB DGX Spark. Hosted, 1 billion input and 100 million output tokens a month is about $260 at OpenRouter's GMICloud price.
- Limits
- Preview release with light post-training; low factual recall (10.1 on SimpleQA-Verified for the base model). The DGX Spark script handles one request at a time with a 131,072-token context and no HTTP server, according to a DEV Community review of the file. The README warns that non-strict tool calls can add extra argument fields.
What’s good
- 80.5 on Terminal-Bench 2.1, level with models up to 15 times larger (self-reported)
- Apache 2.0 weights, including the base model
- 1M-token context on Ricursive's two-B200 server recipe
- Cheap hosted access at $0.20 / $0.60 per 1M tokens
What’s not
- No independent benchmark results yet
- Trails Reflection Beam on SWE-bench Verified, SWE-bench Pro v1.0 and MCP Atlas
- BF16 needs about 130 GB; the DGX Spark script stops at 131,072 tokens
- Needs trust_remote_code and Ricursive's own SGLang parsers
Best for
Platform teams with two B200s, or a DGX Spark, who want an Apache 2.0 coding-agent model they can host and fine-tune themselves.
Ripple-65B Preview alternatives
Ripple-65B Preview: common questions
- Is Ripple-65B Preview worth it?
- Ripple-65B makes a credible case that a 65B dense model can match American open models eight to fifteen times its size at agentic coding, but every score so far is Ricursive's own. We rate it 3.9 out of 5.
- How much does Ripple-65B Preview cost?
- Free open weights under Apache 2.0. Hosted on OpenRouter (GMICloud) at $0.20 per 1M input tokens and $0.60 per 1M output tokens.
- What are the downsides of Ripple-65B Preview?
- No independent benchmark results yet; trails Reflection Beam on SWE-bench Verified, SWE-bench Pro v1.0 and MCP Atlas; bF16 needs about 130 GB; the DGX Spark script stops at 131,072 tokens; needs trust_remote_code and Ricursive's own SGLang parsers.
- Who is Ripple-65B Preview best for?
- Platform teams with two B200s, or a DGX Spark, who want an Apache 2.0 coding-agent model they can host and fine-tune themselves.
- How much does Ripple-65B Preview cost for a team or in production?
- BF16 weights are about 130 GB by our arithmetic from the 64.8 billion parameters, so plan on two B200s; OrcaRouter estimates that pair at about $4,800 a month. The NVFP4 files total 57.2 GB for a 128 GB DGX Spark. Hosted, 1 billion input and 100 million output tokens a month is about $260 at OpenRouter's GMICloud price.
- Is Ripple-65B Preview safe for company data?
- Apache 2.0 for all four checkpoints, with no revenue cap or use restrictions beyond the licence. Self-hosted prompts stay on your hardware; hosted terms are GMICloud's, and OpenRouter offers no EU in-region routing for it.
- What are the best Ripple-65B Preview alternatives?
- The closest alternatives we have reviewed are Beam, Muse Spark Open, Flint and DeepSeek-V4.1-Flash.
Answers checked Nov 5, 2026.