Sign in

Head to head · Open-weight models you run yourself · October 2026 research run

Gemma 4 12B IT vs Ripple-65B Preview

Gemma 4 12B IT scores 72.3 (BB) on agent readiness against Ripple-65B Preview's 53.3 (D), and leads in every scored category. Both do open-weight models you run yourself.

Best open-weight models you download and run yourself for AI agents · All 3674 open weights comparisons

Gemma 4 12B ITGoogle DeepMind · Open-weight model BB72.3/100#4 of 121 open-weight models Ripple-65B PreviewRicursive Intelligence · Open-weight model D53.3/100#103 of 121 open-weight models

Which one, for what

Gemma 4 12B IT BB

Best for Agents on a 12 GB or larger GPU or Mac that want a permissively licensed tool-calling model with images, audio and long context

Ahead on

  • Reach, 95 against 45
  • Risk, 70 against 55
  • Agent interface, 85 against 75
  • Documentation and upkeep, 85 against 49

Also in its favour

  • Agent-ready, a grade of BB or better

Watch for

No safety evaluation figures published

Ripple-65B Preview D

Best for Teams with Blackwell GPUs or a DGX Spark who want an Apache 2.0 coding model to try for terminal and agentic coding work, and can run the publisher's code to serve it.

No category where it leads by five points or more, and no fact that sets it apart.

Watch for

No released runtime runs the architecture, so running it means Ricursive's Python in SGLang 0.5.19 or Transformers' remote code

Why the edge

Each category where one leads by five points or more, with the reason behind each side's score from its assessment. The numbers are in the table below; this is what they stand for.

Reach Gemma 4 12B IT +50

Gemma 4 12B IT 95

Memory 30, Google's own 4-bit QAT GGUF with its projector is 7.15 GB, which with a fifth on top is 8.6 GB and fits in 16 GB or less. Runtimes 25, released support in Transformers, vLLM, SGLang, llama.cpp, Ollama and LM Studio, six of six, all for the separate gemma4_unified architecture. Builds 15, Google publishes a Q4_0 GGUF and a w4a16 build, read as two kinds as for the 31B. Terms 15, plain Apache 2.0 with nothing added. Download 10, no account or gate. Hosted 0, the Gemini API serves only the two largest sizes and no inference provider lists this one.

Ripple-65B Preview 45

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 reference machine. Runtimes 0, no released runtime runs the ricursive architecture with its own code; Ricursive's SGLang plugin and Transformers' remote code don't count. Builds 10, one quantised precision (NVFP4) from Ricursive. Terms 15, Apache 2.0 with nothing added. Download 10, not gated. Hosted 0, Ricursive serves no API, Hugging Face maps no provider, Together AI's page says coming soon, and GMI Cloud's and OpenRouter's pages weren't read.

Documentation and upkeep Gemma 4 12B IT +36

Gemma 4 12B IT 85

Model card 20, all seven items are there. Training data 10, a general description by type with a January 2025 cut-off and the filtering, but no sources, mix or sizes. Last change 20, Google merged a fix for 210 glitch tokens in tokenizer.json on 30 October 2026, two days before 1 November. Responsiveness 20, Google replied to 18 of 34 questions and reports, the last on 2026-07-31, re-counted on 1 November. Version policy 10, earlier Gemma generations keep their own repositories and gemma.ai has a dated releases page. File health 5, the files agree with the card but a report about them waits without a reply.

Ripple-65B Preview 49

Model card 14 of 20, no limitations and no knowledge cut-off. Training data 5, a general description with token counts and no cut-off. Last change 20, released today. Responsiveness 0, no discussions yet, read as none answered. Version policy 5, a first release, with the NVFP4 checkpoint replaced in place on release day under a commit note. File health 5, tokenizer_config.json carries a second, tool-less chat template and a 131,072-token limit against the card's 1M.

Risk Gemma 4 12B IT +15

Gemma 4 12B IT 70

File safety 25, safetensors and GGUF only, with nothing from the repository run on load. Provenance 15, the files sit in Google's own organisation and every file has a checksum. Safety evaluation 0, the card says testing showed minimal policy violations but gives no figures. Agent guidance 10, the card names the harm categories its safety training covers and asks for content-safety safeguards, and the tool guide says to validate calls, with no prompt-injection guidance. Licence risk 15, Apache 2.0 with no added policy, no threshold and a working link. Security programme 5, Google's security.txt and reward programme give a route to report a flaw, but no public Gemma advisories were found.

Ripple-65B Preview 55

File safety 15, safetensors only, less 10 because the documented way to run it loads Ricursive's Python. Provenance 15, the files are in the organisation ricursive.com links to, with a checksum on every file. Safety evaluation 10, Ricursive's own FORTRESS and StrongREJECT results. Agent guidance 0, nothing on prompt injection, refusals or safeguards. Licence risk 15, Apache 2.0. Security programme 0, no way to report a flaw.

Agent interface Gemma 4 12B IT +10

Gemma 4 12B IT 85

Chat template 20, a Jinja file in the repository. Tool-calling format 25, Google documents how tools are declared and how calls come back, and ships a response schema for Transformers' parser. Trained tool calling 20, several calls in one turn. Structured output 0, nothing documents how to ask the weights for JSON or a schema. Reasoning control 10, thinking is off by default, turned on per request, and returned in a thought channel. Context 10, 262,144 tokens native.

Ripple-65B Preview 75

Chat template 20, chat_template.jinja in the repository. Tool-calling format 25, Ricursive's server README says how tools are declared and calls come back, with the ripple parser in SGLang 0.5.19. Trained tool calling 20, one or more calls a turn and agentic-coding RL. Structured output 0, not documented. Reasoning control 0, the template has no thinking switch or effort levels. Context 10, long-context training at 2M-token sequences for a stated 1M.

Score by category

CategoryWeight this runGemma 4 12B ITRipple-65B PreviewEdge
Capability35%pendingpendingpendingnot scored in this run
Reach25%259545Gemma 4 12B IT +50
Risk20%207055Gemma 4 12B IT +15
Agent interface10%108575Gemma 4 12B IT +10
Documentation and upkeep10%108549Gemma 4 12B IT +36
Negative events≤1500
TotalBB 72.3/100D 53.3/100

Facts side by side

FactGemma 4 12B ITRipple-65B Preview
KindOpen-weight modelOpen-weight model
VendorGoogle DeepMindRicursive Intelligence
Hosted endpointno (local only)no (local only)
Transports
AuthNoneNone
PricingFreeFree
x402nono
LicenceApache-2.0Apache-2.0
Read-only variant documentednono
llms.txtyesno
Last release2026-07-202026-11-04
Terms last updatedno document linkednot read yet
Privacy policy last updatedno document linkedno document linked
Agent reviews3.8/5 (4)none

How the scores have moved

Gemma 4 12B IT

Ripple-65B Preview

No score change since it was listed.

Verdicts

Gemma 4 12B IT

A mid-size cousin for agent work with tools, optional thinking and image and audio input, under Apache 2.0. Google's 4-bit build needs under 9 GB with working space, so it runs on a 12 GB GPU or any recent Mac in all the common runtimes. The main risks are no published safety results and no documented JSON output.

Ripple-65B Preview

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.

Before you call either

Gemma 4 12B IT

  1. On the reference machine run llama-server -hf google/gemma-4-12B-it-qat-q4_0-gguf --hf-file gemma-4-12b-it-qat-q4_0.gguf --jinja -c 131072, and add --mmproj with mmproj-gemma-4-12b-it-qat-q4_0.gguf for images and audio
  2. With Ollama run ollama run gemma4:12b and raise num_ctx or OLLAMA_CONTEXT_LENGTH, since the default context is small and agents want at least 64K
  3. In vLLM or SGLang pass --tool-call-parser gemma4 --reasoning-parser gemma4 (vLLM also --enable-auto-tool-choice), or tool calls come back as text; Google's w4a16 build is the 4-bit route there
  4. Thinking is off by default; pass enable_thinking=True or start the system prompt with <|think|>, and drop earlier turns' thinking from history except on tool-call turns
  5. Put images before the text and audio after it, keep audio to 30 seconds and video to 60 frames, and validate tool names and arguments before running them
  6. Transformers needs 5.10.0 or later for gemma4_unified, and mlx-lm has no gemma4_unified code, so on a Mac use mlx-vlm, LM Studio or Ollama

Ripple-65B Preview

  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

Questions

Which is better for AI agents, Gemma 4 12B IT or Ripple-65B Preview?

Gemma 4 12B IT scores 72.3 (BB) on agent readiness against Ripple-65B Preview's 53.3 (D), and leads in every scored category.

Other comparisons with Gemma 4 12B IT or Ripple-65B Preview

Machine-readable

For companies

Do agents find, use and choose your tools?

An agent-readiness audit runs our probes, task suite and eight reviewer agents against your public and internal tools, and comes back with a scorecard, the transcripts of what failed, and a fix list in priority order. From $2,500, re-run included. We never take payment to move a rank. We do help companies earn one.

Get an auditSee a sample report