SYLIN

Gemma

by Google · Model · Language & code models

mixed

Solid small open model for documents and images, but not the leader

Big update fixes 11 issues: tool calling, chat formatting, thinking retention, vision, speed; same model names and weights

AccessOpen weightsNeeds~6.7 GBDownloadabout 6.7 GB

Why we say so

Arena places Gemma 4 third among open weights on documents (1434) and vision (1279), behind Kimi K2.6, MiniMax M3 Flash and GLM-5.3. Liquid AI claims its own LFM 2.5 beats Gemma 4, but that is a maker's claim. MIT authors noted Gemma lacked the looping problem their editing fixed in other models, and no hands-on reviews are on record.

Ranks #3 for documents (open weights) · #3 for image understanding (open weights)

Evidence

  1. Nov 9
    #3 on Arena document among open weights (1434)
  2. Nov 9
    #3 on Arena vision among open weights (1279)
  3. Nov 8
    Weight-space editing by Astra could not improve Gemma, which lacked the repetitive-loop problem
  4. Oct 11
    Gemma 4 used as an agent that plans hypotheses, writes pipelines and runs analysis in the Botanic 1 system
  5. Sep 13
    #3 on Arena document among open weights (1425)
Show all 11
  1. Sep 3
    Gemma 4 31B on Cerebras praised for instant answers and intent understanding in agentic apps, via Kagi
  2. Aug 25
    Commenter asks whether Gemma 4 E4B would be more practical than 3B-active Qwen MoE on a Pi
  3. Aug 18
    Gemma 4 serves as the text encoder for LTX 2.5
  4. Jul 18
    4-bit QAT 12B GGUF from Unsloth is 6.7 GB versus about 24 GB full precision; runs on one GPU
  5. Jul 18
    4-bit QAT via Hermes Agent found the planted port-draft clearance bug (+0.5 instead of -0.5) unprompted and proved the fix
  6. Jul 17
    Configurable vision token budget: choose how many tokens an image gets, from low-cost 70 up to 1120 for OCR and fine text

Cost to run

Alternatives

Kimi#1 on Arena document among open weights (1450)
MiniMax M Flash#2 on Arena document among open weights (1437)
GLM#1 on Arena vision among open weights (1295)
NanbeigeCharted ahead of Gemma 4 and Qwen 3.5 4B and 9B on MCP Atlas, SWE-Bench, Terminal-Bench, HLE and GPQA
Nemotron 3.5 LightningPer NVIDIA's chart, does not win coding or terminal benchmarks against same-size peers; wins instruction following and long context