model
Gemma 4 31B llm
A model record: onboarded once for the project and used by any eval's arms.
- weights
- open
- licence
- Apache-2.0
- aliases
gemma4:31bgemma4:31b-it-q4_K_Meval-gemma-31b- settings
reasoning_effortnone it reasoned by default on the onboarding probe (2026-10-31); off unless a task needs it (RUNBOOK §16): left at the server's default, one model reasons at length and the next not at all- onboarding
- vetted 2026-10-31
- best result per goal
- #2 of 33 A scorer that agrees with experts on summaries (SummEval)
#6 of 19 Evidence retrieval for scientific claims (SciFact)
#9 of 28 Named-entity extraction (Few-NERD)
#13 of 28 Encyclopedia topic classification (DBpedia)
#14 of 30 News topic classification (AG News)
#27 of 55 Triage decisions on Orrery's error signals
#16 of 29 Short news summaries (CNN)
01deployments and their onboarding
x3 · x3 · ollama 0.40.13 · gemma4:31b · onboarding pass 2026-10-31 · 20.1 tokens/s
- where
- x3
- server or device
- ollama 0.40.13
- build
gemma4:31b17ba34c06c801ba22f6047bab21ced9429c6d43e0e42ecf339b0f3e5a61e8d97Q4_K_M- matches the record
- yes
- checked
- 2026-10-31
- onboarded on
- x3
| probe | measurement | flag |
|---|---|---|
| endpoint | provider openai · base url http://127.0.0.1:11434/v1 · model gemma4:31b | — |
| digest found | 17ba34c06c801ba22f6047bab21ced9429c6d43e0e42ecf339b0f3e5a61e8d97 | — |
| default | s 14.7 · tokens out 193 · reasoned: yes · finish stop | — |
| chosen | s 1.15 · tokens out 23 · reasoned: no · finish stop · settings (reasoning effort none) | — |
| tokens per s | 20.1 | — |
| deterministic | yes | — |
| json valid | yes | — |
dgx · dgx · ollama 0.40.13 · gemma4:31b · onboarding pass 2026-10-31 · 23.5 tokens/s
- where
- dgx
- server or device
- ollama 0.40.13
- build
gemma4:31b17ba34c06c801ba22f6047bab21ced9429c6d43e0e42ecf339b0f3e5a61e8d97- matches the record
- yes
- checked
- 2026-10-31
- onboarded on
- dgx
| probe | measurement | flag |
|---|---|---|
| endpoint | provider openai · base url http://100.69.49.126:11434/v1 · model gemma4:31b | — |
| digest found | 17ba34c06c801ba22f6047bab21ced9429c6d43e0e42ecf339b0f3e5a61e8d97 | — |
| default | s 14.42 · tokens out 193 · reasoned: yes · finish stop | — |
| chosen | s 0.98 · tokens out 23 · reasoned: no · finish stop · settings (reasoning effort none) | — |
| tokens per s | 23.5 | — |
| deterministic | yes | — |
| json valid | yes | — |
mac · mac · ollama 0.40.13 · gemma4:31b-it-q4_K_M · onboarding pass 2026-10-31 · 11.5 tokens/s
- where
- mac
- server or device
- ollama 0.40.13
- build
gemma4:31b-it-q4_K_M17ba34c06c801ba22f6047bab21ced9429c6d43e0e42ecf339b0f3e5a61e8d97Q4_K_M- matches the record
- yes
- checked
- 2026-10-31
- onboarded on
- mac
| probe | measurement | flag |
|---|---|---|
| endpoint | provider openai · base url http://127.0.0.1:11435/v1 · model gemma4:31b-it-q4_K_M | — |
| digest found | 17ba34c06c801ba22f6047bab21ced9429c6d43e0e42ecf339b0f3e5a61e8d97 | — |
| default | s 24.35 · tokens out 193 · reasoned: yes · finish stop | — |
| chosen | s 2.0 · tokens out 23 · reasoned: no · finish stop · settings (reasoning effort none) | — |
| tokens per s | 11.5 | — |
| deterministic | yes | — |
| json valid | yes | — |
hosted · homelab LiteLLM (the fleet's gateway) · eval-gemma-31b · onboarding pass 2026-10-31 · 17.9 tokens/s · 2 accepted
- where
- homelab LiteLLM (the fleet's gateway)
- build
eval-gemma-31b- matches the record
- yes
- checked
- 2026-10-31
- onboarded on
- homelab LiteLLM (the fleet's gateway)
| probe | measurement | flag |
|---|---|---|
| endpoint | provider litellm · base url http://homelab:4001/v1 · model eval-gemma-31b | — |
| digest found | None | — |
| default | s 1.41 · tokens out 23 · reasoned: no · finish stop | — |
| chosen | s 1.28 · tokens out 23 · reasoned: no · finish stop · settings (reasoning effort none) | — |
| tokens per s | 17.9 | — |
| deterministic | yes | — |
| json valid | yes | — |
Accepted: where it ran: a hosted gateway serves the provider's current build; there is no digest, revision or hash to pin; determinism: a hosted provider batches requests, so temperature 0 does not repeat exactly; repeats measure it
What it can do, across its deployments: reasons: no · json mode: yes · reasoning off reasoning_effort: none
02settings history
| when | what changed | why |
|---|---|---|
| 2026-10-31 | reasoning_effort: (server default) → none | it reasoned by default on the onboarding probe (2026-10-31); off unless a task needs it (RUNBOOK §16): left at the server's default, one model reasons at length and the next not at all |
03arms