DiffusionGemma 26B A4B IT vs Ling-3.1-flash-1019

DiffusionGemma 26B A4B IT 70

File safety gets the full 25, safetensors only and loaded by Transformers' own code and a Jinja template. Provenance gets 15 for files in Google's organisation with a hash for every file. Licence risk keeps 15 under plain Apache 2.0. Agent guidance gets 10 for the card's list of harm categories and its advice to add content-safety safeguards. Safety evaluation scores 0 because the card states improvements without figures. The security programme gets 5 for Google's security.txt, with no public advisories for Gemma.

Ling-3.1-flash-1019 45

File safety 15, safetensors only, no .bin, .pt or .pth files (25), less 10 because the card's documented vLLM command passes --trust-remote-code, which runs the repository's configuration_bailing_moe_v3.py, since vLLM 0.33.0 registers no bailing_hybrid config of its own. Provenance 15, files in the inclusionAI organisation that inclusion-ai.org links to, per the Ling-3.0-tiny listing (10). Safety evaluation 0, no safety results for the model on its card or in the training-content summary. Agent guidance 0, the card has nothing on prompt injection, refusals or safeguards, and inclusionAI publishes no guardrail probe for this model as it did for Ling 3.0. Licence risk 15, MIT with no use policy, no field-of-use limits, no threshold and no clause letting inclusionAI change or revoke it. Security programme 0, www.inclusion-ai.org answers 404 for security.txt on 10 November 2026, and the Ling-3.0-tiny listing found no SECURITY.md or disclosure policy in the inclusionAI .github repository. 15 + 15 + 0 + 0 + 15 + 0 = 45.

DiffusionGemma 26B A4B IT 70

The chat template is in the repository (20) and handles several tool calls in one turn, and the card names function calling as trained (20). The tool-calling format gets 10, since only the template and the shipped response schema say how tools work here and Google's tool-calling guide is for Gemma 4. Structured output scores 0, undocumented and rejected by SGLang. Thinking can be turned on or off per request with the reasoning in its own channel (10), and the context is 256K (10).

Ling-3.1-flash-1019 60

Chat template 20, chat_template.jinja in the repository (6,042 characters), which handles tools, the thinking switch and preserved reasoning. Tool-calling format 10, only the template says how tools are declared (JSON in <tools>) and how calls come back (<tool_call>name with <arg_key> and <arg_value> pairs). Trained tool calling 20, the template tells the model it may call one or more functions and loops over every entry in an assistant message's tool_calls, and the card reports gains on agentic benchmarks, AA-Briefcase and AutomationBench-AA among them. Structured output 0, the card doesn't document JSON or JSON Schema output. Reasoning control 0, the card documents turning thinking on or off per request with chat_template_kwargs enable_thinking, and names the ling3 reasoning parser, but doesn't say where the reasoning text comes back. Context 10, 262,144 tokens natively in config.json (max_position_embeddings 262144, rope_scaling null) and per SGLang's cookbook. 20 + 10 + 20 + 0 + 0 + 10 = 60.

DiffusionGemma 26B A4B IT 55

Memory earns 15 because the only Google build is 51.7 GB of BF16, about 62 GB with the fifth added, which needs the 64 GB band. Runtimes get 15 for Transformers from 5.11.0, SGLang from v0.5.22 and vLLM, whose releases from v0.24.0 (tag commit 28 June 2026) to v0.33.0 register DiffusionGemmaForBlockDiffusion; llama.cpp (b12114, and Ollama through its pin b11963) and LM Studio don't run it. Builds score 0 because Google publishes no quantised build, though others have published FP8, NVFP4, GGUF and 4-bit copies. Apache 2.0 with nothing added gives terms the full 15 and the ungated repository gives download the full 10. Hosted too scores 0, since no inference provider is mapped and the Gemini API doesn't serve it.

Ling-3.1-flash-1019 60

Memory 0, inclusionAI's only build is the FP8 release at 573.78 GB, × 1.2 = 688.5 GB, above 256 GB, and no runtime's library has a build. Runtimes 25, Released model code in Transformers 5.22.2 (bailing_hybrid), vLLM 0.33.0 (BailingMoeV3ForCausalLM, with ling3 tool and reasoning parsers), SGLang 0.5.23 (bailing_moe_v3.py, its bailing_hybrid config and MTP module) and llama.cpp b12242 (bailingmoe3), and Ollama, credited through its pinned llama.cpp as on the Ling-3.0-tiny listing: v0.40.16 sets LLAMA_CPP_VERSION to b12177. Builds 10, the release is FP8, the one quantised precision inclusionAI publishes for it. Terms 15, MIT in the card's metadata, with no use policy or other terms added on the card or in the repository. Download 10, the Hub API's model info gives gated false on 10 November 2026, so the files download with no account and no terms to accept. Hosted 0, the Hub API's provider mapping is empty for this repository on 10 November 2026. 0 + 25 + 10 + 15 + 10 + 0 = 60.