Gemma 4 26B
An open-source large language model in Google's Gemma series with 26B parameters, suited to inference and fine-tuning tasks.
Timeline (last 90 days)
In a test of 360 clean texts, Jev made only 1 misjudgment, gpt-5-mini made 1 misjudgment, Gemini 2.5 Flash-Lite and Gemma each made 2 misjudgments, gpt-oss-safeguard-20b made 4 misjudgments, and Llama Guard 4 made 26 misjudgments.
Running Gemma 4 26B on an M3 Max, MLX is 65.0 tokens/s and LLAMA.cpp is 63.9 tokens/s, essentially a tie between the two.
Based on public data for high power plus MLX or Ollama, the decoding speed of different models on M4 Max is roughly a ladder as follows: Nemotron 3 Nano about 160 tokens/s, Qwen 3.5 35B at 107 to 130 tokens/s, Gemma 4 26B about 75 tokens/s, DeepSeek V4 Flash about 27 tokens/s, and dense 27B about 22 tokens/s.
On a Korean NPU server (8 cards, 32 chips in total), VKORE raises the single-user response of the Gemma 4 26B model from 14.1 tokens/s to 53.9 tokens/s, about 3.8 times.
Qwen 3 235B and Gemma 4 26B both scored 0.6875 on resolve-plus, the lowest in the whole field.
Gemma 4 26B A4B has 25.2B total parameters / 3.8B active; Qwen3-Next-80B-A3B has 80B / 3B; gpt-oss-20b has 20.9B / 3.6B.
Gemma 4 26B terminated with a runtime error before producing a score, because the scoring code assumed at least one sentence would be returned in the correct structured format.
The demo runs entirely on a local environment using the Gemma 4 26B model.
All Facts (8)
In a test of 360 clean texts, Jev made only 1 misjudgment, gpt-5-mini made 1 misjudgment, Gemini 2.5 Flash-Lite and Gemma each made 2 misjudgments, gpt-oss-safeguard-20b made 4 misjudgments, and Llama Guard 4 made 26 misjudgments.
50%UnverifiedRunning Gemma 4 26B on an M3 Max, MLX is 65.0 tokens/s and LLAMA.cpp is 63.9 tokens/s, essentially a tie between the two.
50%UnverifiedBased on public data for high power plus MLX or Ollama, the decoding speed of different models on M4 Max is roughly a ladder as follows: Nemotron 3 Nano about 160 tokens/s, Qwen 3.5 35B at 107 to 130 tokens/s, Gemma 4 26B about 75 tokens/s, DeepSeek V4 Flash about 27 tokens/s, and dense 27B about 22 tokens/s.
50%UnverifiedOn a Korean NPU server (8 cards, 32 chips in total), VKORE raises the single-user response of the Gemma 4 26B model from 14.1 tokens/s to 53.9 tokens/s, about 3.8 times.
50%UnverifiedQwen 3 235B and Gemma 4 26B both scored 0.6875 on resolve-plus, the lowest in the whole field.
50%UnverifiedGemma 4 26B A4B has 25.2B total parameters / 3.8B active; Qwen3-Next-80B-A3B has 80B / 3B; gpt-oss-20b has 20.9B / 3.6B.
50%UnverifiedThe demo runs entirely on a local environment using the Gemma 4 26B model.
50%UnverifiedGemma 4 26B terminated with a runtime error before producing a score, because the scoring code assumed at least one sentence would be returned in the correct structured format.
50%