model family

gemma-4

10 members have been measured on real machines. Every number below was signed by the machine that produced it — none of it is from a model card.

what each build proved
e2b · Q4_K_M · ollama
e2b · Q3_K_M
e2b · Q4_K_M · unsloth
e2b · IQ4_NL
e2b · Q4_0
e2b · Q3_K_S
e4b · Q3_K_M
e2b · e2b-it-qat
e4b · e4b-it-qat
e2b
e2b · Q4_K_M · ollama · Tool calling — provede2b · Q4_K_M · ollama · Tool loop finished — provede2b · Q4_K_M · ollama · Faithful to tool results — provede2b · Q4_K_M · ollama · Chained tool calls — provede2b · Q4_K_M · ollama · Structured output — provede2b · Q4_K_M · ollama · Sustained output — provede2b · Q4_K_M · ollama · Extended thinking — provede2b · Q4_K_M · ollama · Vision — provede2b · Q4_K_M · ollama · Counts in images — provede2b · Q4_K_M · ollama · Multiple images at once — provede2b · Q4_K_M · ollama · Spatial reasoning — provede2b · Q4_K_M · ollama · Writes code — provede2b · Q4_K_M · ollama · Reasons about code — provede2b · Q4_K_M · ollama · Synthesises sources — provede2b · Q4_K_M · ollama · Follows instructions — provede2b · Q3_K_M · Tool calling — provede2b · Q3_K_M · Tool loop finished — provede2b · Q3_K_M · Faithful to tool results — provede2b · Q3_K_M · Chained tool calls — provede2b · Q3_K_M · Structured output — provede2b · Q3_K_M · Sustained output — provede2b · Q3_K_M · Extended thinking — provede2b · Q3_K_M · Vision — provede2b · Q3_K_M · Counts in images — provede2b · Q3_K_M · Multiple images at once — provede2b · Q3_K_M · Spatial reasoning — provede2b · Q3_K_M · Writes code — provede2b · Q3_K_M · Reasons about code — provede2b · Q3_K_M · Synthesises sources — provede2b · Q3_K_M · Follows instructions — provede2b · Q4_K_M · unsloth · Tool calling — provede2b · Q4_K_M · unsloth · Tool loop finished — provede2b · Q4_K_M · unsloth · Faithful to tool results — provede2b · Q4_K_M · unsloth · Chained tool calls — provede2b · Q4_K_M · unsloth · Structured output — provede2b · Q4_K_M · unsloth · Sustained output — provede2b · Q4_K_M · unsloth · Extended thinking — provede2b · Q4_K_M · unsloth · Vision — provede2b · Q4_K_M · unsloth · Counts in images — provede2b · Q4_K_M · unsloth · Multiple images at once — provede2b · Q4_K_M · unsloth · Spatial reasoning — provede2b · Q4_K_M · unsloth · Writes code — provede2b · Q4_K_M · unsloth · Reasons about code — provede2b · Q4_K_M · unsloth · Synthesises sources — provede2b · Q4_K_M · unsloth · Follows instructions — provede2b · IQ4_NL · Tool calling — provede2b · IQ4_NL · Tool loop finished — provede2b · IQ4_NL · Faithful to tool results — provede2b · IQ4_NL · Chained tool calls — provede2b · IQ4_NL · Structured output — provede2b · IQ4_NL · Sustained output — provede2b · IQ4_NL · Extended thinking — provede2b · IQ4_NL · Vision — provede2b · IQ4_NL · Counts in images — provede2b · IQ4_NL · Multiple images at once — provede2b · IQ4_NL · Spatial reasoning — provede2b · IQ4_NL · Writes code — provede2b · IQ4_NL · Reasons about code — provede2b · IQ4_NL · Synthesises sources — provede2b · IQ4_NL · Follows instructions — provede2b · Q4_0 · Tool calling — provede2b · Q4_0 · Tool loop finished — provede2b · Q4_0 · Faithful to tool results — provede2b · Q4_0 · Chained tool calls — provede2b · Q4_0 · Structured output — provede2b · Q4_0 · Sustained output — provede2b · Q4_0 · Extended thinking — provede2b · Q4_0 · Vision — provede2b · Q4_0 · Counts in images — provede2b · Q4_0 · Multiple images at once — provede2b · Q4_0 · Spatial reasoning — provede2b · Q4_0 · Writes code — provede2b · Q4_0 · Reasons about code — provede2b · Q4_0 · Synthesises sources — provede2b · Q4_0 · Follows instructions — provede2b · Q3_K_S · Tool calling — provede2b · Q3_K_S · Tool loop finished — asked, failede2b · Q3_K_S · Faithful to tool results — provede2b · Q3_K_S · Chained tool calls — provede2b · Q3_K_S · Structured output — provede2b · Q3_K_S · Sustained output — provede2b · Q3_K_S · Extended thinking — provede2b · Q3_K_S · Vision — provede2b · Q3_K_S · Counts in images — provede2b · Q3_K_S · Multiple images at once — provede2b · Q3_K_S · Spatial reasoning — provede2b · Q3_K_S · Writes code — provede2b · Q3_K_S · Reasons about code — provede2b · Q3_K_S · Synthesises sources — provede2b · Q3_K_S · Follows instructions — provede4b · Q3_K_M · Tool calling — provede4b · Q3_K_M · Tool loop finished — provede4b · Q3_K_M · Faithful to tool results — provede4b · Q3_K_M · Chained tool calls — provede4b · Q3_K_M · Structured output — provede4b · Q3_K_M · Sustained output — provede4b · Q3_K_M · Extended thinking — provede4b · Q3_K_M · Vision — provede4b · Q3_K_M · Counts in images — provede4b · Q3_K_M · Multiple images at once — provede4b · Q3_K_M · Spatial reasoning — provede4b · Q3_K_M · Writes code — provede4b · Q3_K_M · Reasons about code — provede4b · Q3_K_M · Synthesises sources — provede4b · Q3_K_M · Follows instructions — provede2b · e2b-it-qat · Tool calling — never askede2b · e2b-it-qat · Tool loop finished — never askede2b · e2b-it-qat · Faithful to tool results — never askede2b · e2b-it-qat · Chained tool calls — never askede2b · e2b-it-qat · Structured output — never askede2b · e2b-it-qat · Sustained output — never askede2b · e2b-it-qat · Extended thinking — never askede2b · e2b-it-qat · Vision — never askede2b · e2b-it-qat · Counts in images — never askede2b · e2b-it-qat · Multiple images at once — never askede2b · e2b-it-qat · Spatial reasoning — never askede2b · e2b-it-qat · Writes code — never askede2b · e2b-it-qat · Reasons about code — never askede2b · e2b-it-qat · Synthesises sources — never askede2b · e2b-it-qat · Follows instructions — never askede4b · e4b-it-qat · Tool calling — never askede4b · e4b-it-qat · Tool loop finished — never askede4b · e4b-it-qat · Faithful to tool results — never askede4b · e4b-it-qat · Chained tool calls — never askede4b · e4b-it-qat · Structured output — never askede4b · e4b-it-qat · Sustained output — never askede4b · e4b-it-qat · Extended thinking — never askede4b · e4b-it-qat · Vision — never askede4b · e4b-it-qat · Counts in images — never askede4b · e4b-it-qat · Multiple images at once — never askede4b · e4b-it-qat · Spatial reasoning — never askede4b · e4b-it-qat · Writes code — never askede4b · e4b-it-qat · Reasons about code — never askede4b · e4b-it-qat · Synthesises sources — never askede4b · e4b-it-qat · Follows instructions — never askede2b · Tool calling — never askede2b · Tool loop finished — never askede2b · Faithful to tool results — never askede2b · Chained tool calls — never askede2b · Structured output — never askede2b · Sustained output — never askede2b · Extended thinking — never askede2b · Vision — never askede2b · Counts in images — never askede2b · Multiple images at once — never askede2b · Spatial reasoning — never askede2b · Writes code — never askede2b · Reasons about code — never askede2b · Synthesises sources — never askede2b · Follows instructions — never asked
Tool calling
Tool loop finished
Faithful to tool results
Chained tool calls
Structured output
Sustained output
Extended thinking
Vision
Counts in images
Multiple images at once
Spatial reasoning
Writes code
Reasons about code
Synthesises sources
Follows instructions
gemma4:e2be2b · Q4_K_M · 5 machines · 5.1B
8,192recalled
32,768served
45.1best tok/s
43.3 tok/s · unnamed machine19.3 tok/s · unnamed machine8.9 tok/s · unnamed machine7.7 tok/s · x86 · CPU · Hyderabad4.1 tok/s · unnamed machine
4.1 → 43.3 tok/s across 5 machines · ×10.7 spreadunnamed machinex86 · CPU
recalled 8k▲ served 32k⌐ declared 128k · unsignedadvertises 16× what it proved
hf.co/unsloth/gemma-4-E2B-it-GGUF:Q3_K_Me2b · Q3_K_M · 3 machines · 4.65B
8,192recalled
—served
38.4best tok/s
38.4 tok/s · unnamed machine14.2 tok/s · x86 · CPU · Hyderabad2.8 tok/s · unnamed machine
2.8 → 38.4 tok/s across 3 machines · ×13.8 spreadunnamed machinex86 · CPU
recalled 8k⌐ declared 128k · unsignedadvertises 16× what it proved
hf.co/unsloth/gemma-4-E2B-it-GGUF:Q4_K_Me2b · Q4_K_M · 3 machines · 4.65B
8,192recalled
8,192served
164.7best tok/s
54.7 tok/s · unnamed machine6.8 tok/s · x86 · CPU · Hyderabad3.9 tok/s · unnamed machine
3.9 → 54.7 tok/s across 3 machines · ×14.2 spreadunnamed machinex86 · CPU
recalled 8k⌐ declared 128k · unsignedadvertises 16× what it proved
hf.co/unsloth/gemma-4-E2B-it-GGUF:IQ4_NLe2b · IQ4_NL · 2 machines · 4.65B
8,192recalled
—served
54.9best tok/s
54.9 tok/s · unnamed machine10.0 tok/s · x86 · CPU · Hyderabad
10.0 → 54.9 tok/s across 2 machines · ×5.5 spreadunnamed machinex86 · CPU
recalled 8k⌐ declared 128k · unsignedadvertises 16× what it proved
hf.co/unsloth/gemma-4-E2B-it-GGUF:Q4_0e2b · Q4_0 · 2 machines · 4.65B
4,096recalled
—served
57.6best tok/s
57.6 tok/s · unnamed machine9.1 tok/s · x86 · CPU · Hyderabad
9.1 → 57.6 tok/s across 2 machines · ×6.3 spreadunnamed machinex86 · CPU
recalled 4k⌐ declared 128k · unsignedadvertises 32× what it proved
hf.co/unsloth/gemma-4-E2B-it-GGUF:Q3_K_Se2b · Q3_K_S · 1 machine · 4.65B
4,096recalled
—served
14.4best tok/s
recalled 4k⌐ declared 128k · unsignedadvertises 32× what it proved
hf.co/unsloth/gemma-4-E4B-it-GGUF:Q3_K_Me4b · Q3_K_M · 1 machine
4,096recalled
—served
20.4best tok/s
gemma4:e2b-it-qate2b · e2b-it-qat · 1 machine
—recalled
—served
8.9best tok/s
gemma4:e4b-it-qate4b · e4b-it-qat · 1 machine
—recalled
—served
—best tok/s
google/gemma-4-E2B-it-qat-q4_0-ggufe2b · 1 machine
—recalled
—served
2.0best tok/s
ask your own gram

Your concierge can answer this from the network rather than from this page. The ask carries the operator id, so it resolves to a capability and an engine — not to a name somebody has to recognise.

What has this network measured about the gemma-4 family (epn.family.gemma-4), and can my gram run any of it? See /inferences/gemma-4

open a gramx with this ask →epn.family.gemma-4