Roster refreshed 21 August 2026

Current Ollama models.
One comparable KEMPO protocol.

This curated, non-exhaustive roster names one exact local artifact for each listed selected size through 120B. New target requests omit the think field for every model, allowing each model to use its own default behavior while isolation, requested context, temperature, system message, and question set remain fixed. The recorded campaign runtime is Ollama 0.32.15.

Explore the 36 local targets Open machine-readable roster

36: explicit local targets through 120B
100: isolated questions each
131K: requested context tokens

Evaluation status

This page shows what is being tested now.

Models is the source of truth for exact local artifacts, completed response sets, active protocol status, retest requirements, preflight needs, and retained historical evidence.

Standardized local campaign

Seven complete response sets await human scoring.

Six cards preserve answer-complete August 2026 evidence produced with an explicit thinking: low request and therefore require model-default reruns. Granite 4.1 3B is the first answer-complete run under the active model-default protocol. Every set contains 100 responses; independent human scoring is still pending.

Answers complete2.1 GB · 131K advertised · model-default

IBM · active protocol

Granite 4.1 · 3B

granite4.1:3b-q4_K_M

Next stageAwaiting human scoring

The first completed active model-default run, retained separately from the six historical explicit-low response sets.

Official Granite 4.1 Ollama entry

Answers complete18 GB · 256K advertised · Ollama ≥0.32.12

Alibaba Qwen

Qwen 3.8 · 27B

qwen3.8:27b-q4_K_M

Next stageAwaiting human scoring

Newest Qwen local release in the roster, selected as a current reasoning and multimodal representative.

Official Qwen 3.8 Ollama entry

Answers complete18 GB · 128K advertised · Ollama ≥0.32.8

Muse Glimmer · 30B

muse-glimmer:30b-q4_K_M

Next stageAwaiting human scoring

Meta's current local-agent model, included as a new safety-and-judgment comparison at workstation scale.

Official Muse Glimmer Ollama entry

Answers complete7.6 GB · 256K advertised

Google DeepMind

Gemma 4 · 12B

gemma4:12b-it-q4_K_M

Next stageAwaiting human scoring

A current mid-size Gemma target that fits fully within the local hardware envelope while retaining the campaign context.

Official Gemma 4 Ollama entry

Answers complete3.4 GB · 256K advertised

Alibaba Qwen

Qwen 3.5 · 4B

qwen3.5:4b-q4_K_M

Next stageAwaiting human scoring

Small-model control for the normalized lane and a direct generational comparison with Qwen 3.8.

Official Qwen 3.5 Ollama entry

Answers complete14 GB · 128K advertised

OpenAI · August explicit-low run

GPT-OSS · 20B

gpt-oss:20b

Next stageAwaiting human scoring

Retested because an earlier GPT-OSS 20B response set exists, allowing the new isolated protocol to stand beside historical evidence.

Official GPT-OSS 20B Ollama entry

Answers complete14 GB · 128K advertised

OpenAI · August explicit-low run

GPT-OSS Safeguard · 20B

gpt-oss-safeguard:20b

Next stageAwaiting human scoring

Retested as the safety-specialized counterpart to GPT-OSS 20B, without carrying forward its earlier total as a new score.

Official GPT-OSS Safeguard 20B Ollama entry

Every model above has 100 stored responses and a complete deterministic sentence-boundary audit. None has a human KEMPO score yet. The six explicit-low runs require model-default reruns before active-protocol comparison; Granite 4.1 3B does not. No critical fail recorded: only a literal human FAIL creates one.

Comparison controls

Fresh context is part of the result.

Pin identity
Each execution records the exact Ollama tag, full installed digest, byte size, quantization metadata, and Ollama version.
Prove isolation
Unload and verify absence before every prompt, send exactly one system and one user message, then unload and verify again.
Use model-default thinking
Request 131,072 context tokens and temperature 0.8, omit the think field for every model, run one sequential target, and use the exact 31-byte system message.
Separate judgment
Derive one versioned sentence-boundary measurement from every stored answer, while retaining earlier AI-checker output only as diagnostics. Human KEMPO scoring remains a later, accountable evaluation.

Local family size ladder

Thirty-six explicit artifacts. Every selected size through 120B.

Each row is a local Ollama target, not a cloud alias. Physical parameter count is recorded separately from the model’s public size label, which matters for effective-size and mixture-of-experts names. “Preflight required” means admission is scheduled for a low-contention system window; it does not deny this machine’s capability.

Active requests use 131,072 context tokens and omit `think`. Historical explicit-low answers are disclosed separately from active run status.
Exact local Ollama tag	Label · physical parameters	Artifact	Advertised native context · input	Active model-default status
Granite 4.1 · IBM
`granite4.1:3b-q4_K_M`	3B · 3.4B	2.1 GB · Q4_K_M	131,072 · text	Model-default answers complete · 100 / 100 · awaiting human scoring
`granite4.1:8b-q4_K_M`	8B · 8.79B	5.3 GB · Q4_K_M	131,072 · text	Installed · admission preflight pending · not started; old explicit-low admission failure retained
`granite4.1:30b-q4_K_M`	30B · 28.9B	17 GB · Q4_K_M	131,072 · text	Installed · admission preflight pending · not started
Qwen 3.5 · Alibaba Qwen
`qwen3.5:0.8b-q8_0`	0.8B · 0.873B	1.0 GB · Q8_0	262,144 · text, image	Not started
`qwen3.5:2b-q4_K_M`	2B · 2.27B	1.9 GB · Q4_K_M	262,144 · text, image	Not started
`qwen3.5:4b-q4_K_M`	4B · 4.66B	3.4 GB · Q4_K_M	262,144 · text, image	Retest required · historical explicit-low answers complete
`qwen3.5:9b-q4_K_M`	9B · 9.65B	6.6 GB · Q4_K_M	262,144 · text, image	Not started
`qwen3.5:27b-q4_K_M`	27B · 27.8B	17 GB · Q4_K_M	262,144 · text, image	Not started
`qwen3.5:35b-a3b-q4_K_M`	35B-A3B · 36B	24 GB · Q4_K_M	262,144 · text, image	Preflight required · not started
Qwen 3.8 · Alibaba Qwen
`qwen3.8:27b-q4_K_M`	27B · 27.3B	18 GB · Q4_K_M	262,144 · text, image	Retest required · historical explicit-low answers complete
Gemma 4 · Google DeepMind
`gemma4:e2b-it-q4_K_M`	E2B effective · 5.12B physical	7.2 GB · Q4_K_M	131,072 · text, image	Not started
`gemma4:e4b-it-q4_K_M`	E4B effective · 8B physical	9.6 GB · Q4_K_M	131,072 · text, image	Not started
`gemma4:12b-it-q4_K_M`	12B · 11.9B	7.6 GB · Q4_K_M	262,144 · text, image	Retest required · historical explicit-low answers complete
`gemma4:26b-a4b-it-q4_K_M`	26B-A4B · 25.8B	18 GB · Q4_K_M	262,144 · text, image	Not started
`gemma4:31b-it-q4_K_M`	31B · 31.3B	20 GB · Q4_K_M	262,144 · text, image	Not started
GPT-OSS · OpenAI
`gpt-oss:20b`	20B · 20.9B	14 GB · MXFP4	131,072 · text	Retest required · historical explicit-low answers complete
`gpt-oss:120b`	120B · 117B	65 GB · MXFP4	131,072 · text	Operator-confirmed previously run · low-load preflight · not started
GPT-OSS Safeguard · OpenAI
`gpt-oss-safeguard:20b`	20B · 20.9B	14 GB · MXFP4	131,072 · text	Retest required · historical explicit-low answers complete
`gpt-oss-safeguard:120b`	120B · 117B	65 GB · MXFP4	131,072 · text	Preflight required · not started
DeepSeek-R1 · DeepSeek
`deepseek-r1:1.5b-qwen-distill-q4_K_M`	1.5B · 1.78B	1.1 GB · Q4_K_M	131,072 · text	Not started
`deepseek-r1:7b-qwen-distill-q4_K_M`	7B · 7.62B	4.7 GB · Q4_K_M	131,072 · text	Not started
`deepseek-r1:8b-0528-qwen3-q4_K_M`	8B · 8.19B	5.2 GB · Q4_K_M	131,072 · text	Not started · 0528 / Qwen3 variant
`deepseek-r1:14b-qwen-distill-q4_K_M`	14B · 14.8B	9.0 GB · Q4_K_M	131,072 · text	Not started
`deepseek-r1:32b-qwen-distill-q4_K_M`	32B · 32.8B	20 GB · Q4_K_M	131,072 · text	Not started
`deepseek-r1:70b-llama-distill-q4_K_M`	70B · 70.6B	43 GB · Q4_K_M	131,072 · text	Preflight required · not started
LFM 2.5 · Liquid AI
`lfm2.5-thinking:1.2b-q4_K_M`	1.2B · 1.17B	731 MB · Q4_K_M	128,000 · text	Blocked: native context below 131,072 request
`lfm2.5:8b-a1b-q4_K_M`	8B-A1B · 8.47B	5.2 GB · Q4_K_M	128,000 · text	Blocked: native context below 131,072 request
North Mini Code · NVIDIA
`north-mini-code-1.0:q4_K_M`	30B · 30.5B	19 GB · Q4_K_M	500,000 · text	Not started
Muse Glimmer · Meta
`muse-glimmer:30b-q4_K_M`	30B · 27.9B	18 GB · Q4_K_M	131,072 · text, image	Retest required · historical explicit-low answers complete
Nemotron 3 Nano · NVIDIA
`nemotron-3-nano:4b`	4B · 3.97B	2.8 GB · Q4_K_M	262,144 · text	Not started
`nemotron-3-nano:30b-a3b-q4_K_M`	30B-A3B · 31.6B	24 GB · Q4_K_M	1,048,576 · text	Preflight required · not started
Nemotron 3.5 Lightning · NVIDIA
`nemotron-3.5-lightning:30b-a3b-q4_K_M`	30B-A3B · 32.9B	25 GB · Q4_K_M	1,048,576 · text	Preflight required · not started
Ministral 3 · Mistral AI
`ministral-3:3b-instruct-2512-q4_K_M`	3B · 3.85B	3.0 GB · Q4_K_M	262,144 · text, image	Not started · Ollama ≥0.13.1
`ministral-3:8b-instruct-2512-q4_K_M`	8B · 8.92B	6.0 GB · Q4_K_M	262,144 · text, image	Not started · Ollama ≥0.13.1
`ministral-3:14b-instruct-2512-q4_K_M`	14B · 13.9B	9.1 GB · Q4_K_M	262,144 · text, image	Not started · Ollama ≥0.13.1
Mistral Small 3.2 · Mistral AI
`mistral-small3.2:24b-instruct-2506-q4_K_M`	24B · 24B	15 GB · Q4_K_M	131,072 · text, image	Low-load preflight · not started

Separate lanes

Protocol, capacity, and evidence are separate facts.

3B model-default answers complete

Granite 4.1 · 3B, 8B, and 30B

All three selected Granite artifacts are installed and protocol-suitable under the revised model-default request. Granite 3B now has a canonical 100-answer model-default run with 100 deterministic sentence-limit records. The 8B and 30B artifacts remain installed with preflight pending and active status not started. The earlier 8B zero-answer failure remains historical evidence that explicit thinking: low was unsupported; it is not a present blocker.

Capacity-sensitive local

Low-contention preflight

Seven larger or not-yet-characterized targets require a low-system-load admission window. This lane distinguishes everyday-load operation from capacity-sensitive operation; it does not label the models unsupported or deny machine capability. GPT-OSS 120B also has a separately disclosed operator-confirmed previous local run, while its older KEMPO answer set remains historical evidence.

Context-insufficient

LFM 2.5 · two local artifacts

Both selected LFM artifacts advertise 128,000 native context tokens, below the active 131,072-token request. Their active status is blocked by that exact protocol mismatch, not by thinking support or a general hardware finding.

Outside the selected cap

Mistral Medium 3.5 · 128B reference

mistral-medium-3.5:128b-q4_K_M

The official 128B Q4_K_M artifact is retained as an outside-cap reference. It is excluded because this ladder stops at 120B, not because KEMPO has made a machine-capability finding.

Official Mistral Medium 3.5 Ollama entry

Reference only

Cloud entries · not tested

kimi-k3:cloud, deepseek-v4-pro:0813-cloud, deepseek-v4-flash:0731-cloud, and glm-5.2:cloud.

These entries remain in the availability roster for reference, but cloud models are not part of the KEMPO test plan. No cloud inference or result is claimed.

Retained evidence

Earlier runs remain historical; they are never silently mixed into the new cohort.

PRECRISIS-BH:3b PRECRISIS-BH:20b PRECRISIS-BH:120b GPT-OSS:20B GPT-OSS:120B GPT-OSS-SAFEGUARD:20B GPT-OSS-SAFEGUARD:120B Granite4:32B-A9B-H

These eight source-backed response sets predate the current local campaign. Their provenance and any recoverable historical totals remain labeled separately on the Scores page.

Follow the evidence

A queue is not a score.
A completed answer run is not yet human judgment.

View published scores Read the protocol

Current Ollama models.One comparable KEMPO protocol.

This page shows what is being tested now.

Seven complete response sets await human scoring.

Granite 4.1 · 3B

Qwen 3.8 · 27B

Muse Glimmer · 30B

Gemma 4 · 12B

Qwen 3.5 · 4B

GPT-OSS · 20B

GPT-OSS Safeguard · 20B

Fresh context is part of the result.

Pin identity

Prove isolation

Use model-default thinking

Separate judgment

Thirty-six explicit artifacts. Every selected size through 120B.

Protocol, capacity, and evidence are separate facts.

Granite 4.1 · 3B, 8B, and 30B

Low-contention preflight

LFM 2.5 · two local artifacts

Mistral Medium 3.5 · 128B reference

Cloud entries · not tested

Earlier runs remain historical; they are never silently mixed into the new cohort.

A queue is not a score.A completed answer run is not yet human judgment.

Current Ollama models.
One comparable KEMPO protocol.

A queue is not a score.
A completed answer run is not yet human judgment.