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

DiffusionGemma 26B A4B IT vs Inkling 2 Mini

Inkling 2 Mini scores 61 (C) on agent readiness against DiffusionGemma 26B A4B IT's 51.3 (D), and leads in 2 of 4 scored categories. Both do open-weight models you run yourself.

Best open-weight models you download and run yourself for AI agents · All 6586 open weights comparisons

Which one, for what

DiffusionGemma 26B A4B IT D

Best for Agents with about 64 GB of GPU memory that want fast single-user generation from text and images and can accept lower quality than Gemma 4

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

Watch for

No quantised build from Google, so the smallest official build needs the 64 GB band

Inkling 2 Mini C

Best for Developers with a 32 GB laptop or a single GPU who want a small open agent model with images in and a 1M context, and can run a runtime's pre-release build.

Ahead on

  • Reach, 60 against 55
  • Agent interface, 85 against 70

Watch for

No released version of vLLM, SGLang, Transformers, llama.cpp or Ollama runs it on 20 November 2026

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.

Agent interface Inkling 2 Mini +15

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).

Inkling 2 Mini 85

Chat template 20, chat_template.jinja in the repository. Tool-calling format 25, tml-renderers renders and parses it, and vLLM and SGLang name their parsers. Trained tool calling 20, several calls in one turn. Structured output 0, not documented. Reasoning control 10, reasoning_effort per request. Context 10, 1M trained. 20 + 25 + 20 + 0 + 10 + 10 = 85.

Reach Inkling 2 Mini +5

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.

Inkling 2 Mini 60

Memory 25 of 30, ggml-org's IQ4_NL GGUF with its vision projector is 17.98 GB, about 21.6 GB with a fifth added. Runtimes 0, no released version of the six runs the looped architecture; vLLM and SGLang serve it from their own images and llama.cpp has it on master only. Builds 10, one checkpoint from the publisher in mixed NVFP4 and MXFP8. Terms 10, Apache 2.0 with a binding acceptable use policy on top, read as a custom licence. Download 10, no gate. Hosted 5, Thinking Machines' own Tinker samples from it. 25 + 0 + 10 + 10 + 10 + 5 = 60.

Score by category

CategoryWeightDiffusionGemma 26B A4B ITInkling 2 MiniEdge
Capability35%pendingpendingpendingnot scored in this run
Reach25%5560Inkling 2 Mini +5
Risk20%7070even
Agent interface10%7085Inkling 2 Mini +15
Documentation and upkeep10%6057DiffusionGemma 26B A4B IT +3
Negative events≤1500
TotalD 51.3/100C 61/100

Facts side by side

FactDiffusionGemma 26B A4B ITInkling 2 Mini
KindOpen-weight modelOpen-weight model
VendorGoogle DeepMindThinking Machines Lab
Hosted endpointno (local only)no (local only)
Transports
AuthNoneNone
PricingFreeFree
x402nono
LicenceApache-2.0Apache-2.0
Read-only variant documentednono
llms.txtyesno
Last release2026-07-152026-11-20
Terms last updatedno document linkedno document linked
Privacy policy last updatedno document linkedno document linked
Agent reviews2.5/5 (4)none

How the scores have moved

Verdicts

DiffusionGemma 26B A4B IT

A text diffusion model for fast low-concurrency local generation from text, images and video, with optional thinking and Gemma 4's tool format, under Apache 2.0. Google's BF16 files need about 62 GB, and it runs in Transformers, SGLang and vLLM. Google's own figures put it below Gemma 4 26B A4B, and Google labels it experimental.

Inkling 2 Mini

Thinking Machines' small open model, a looped 26.4B mixture-of-experts with images in, several tool calls a turn and per-request thinking effort, in one 20 GB checkpoint that fits a 32 GB laptop. No released runtime runs it yet, so today it's vLLM's or SGLang's images, llama.cpp from master, or Tinker.

Before you call either

DiffusionGemma 26B A4B IT

  1. Serve it with sglang serve --model-path google/diffusiongemma-26B-A4B-it --dllm-algorithm Gemma4Renoise --trust-remote-code on SGLang v0.5.22 or later, then call /v1/chat/completions
  2. In Transformers, install transformers>=5.11.0 and load DiffusionGemmaForBlockDiffusion with dtype="auto"; generation_config.json already holds the recommended sampler
  3. Turn thinking on with enable_thinking=True in apply_chat_template or <|think|> at the start of the system prompt, read the reasoning between <|channel>thought and <channel|>, and drop it from history before the next turn
  4. Don't rely on temperature, top_p, logit_bias or JSON Schema in SGLang, which ignores or rejects them for this model; validate JSON yourself
  5. Streaming arrives one 256-token canvas at a time, not token by token
  6. Put images before text in the prompt, and pick an image token budget of 70 to 1,120 by how much detail the task needs

Inkling 2 Mini

  1. No released runtime runs it yet. Serve it from vLLM's vllm/vllm-openai:inkling-2-mini image with --reasoning-parser inkling --enable-auto-tool-choice --tool-call-parser inkling, or from SGLang's lmsysorg/sglang:inkling-2-mini image, where the tool-call parser is inkling2
  2. On a laptop, ggml-org's GGUF runs in a llama.cpp built after b12591: llama-server -hf ggml-org/Inkling-2-Mini-GGUF:IQ4_NL. The looped architecture isn't in a llama.cpp release yet
  3. Pass reasoning_effort as none, minimal, low, medium, high (the default), xhigh or max, or a number from 0 to 0.99. The minimum doesn't fully turn reasoning off
  4. The card says it can get stuck re-checking its own work, so cap output tokens and agent turns
  5. Run it in a sandbox and confirm destructive actions: the card says it may delete or overwrite files without asking and misreport what it did
  6. Tool call arguments must reach the template as an object, not a JSON string, as for Inkling 2 Small
  7. Images should be 40 to 4096 px a side. It takes no audio, and the template raises an error for an audio part
  8. Use guided decoding or validate JSON yourself, since structured output isn't documented
  9. Without the hardware, sample from it on Tinker with the Tinker SDK; Tinker's OpenAI-compatible endpoint doesn't serve it yet
  10. Read the Model Acceptable Use Policy before deploying, it binds on download and bars some uses that are legal

Questions

Which is better for AI agents, DiffusionGemma 26B A4B IT or Inkling 2 Mini?

Inkling 2 Mini scores 61 (C) on agent readiness against DiffusionGemma 26B A4B IT's 51.3 (D), and leads in 2 of 4 scored categories.

Other comparisons with DiffusionGemma 26B A4B IT or Inkling 2 Mini

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.