WUMBOLABS / EVALUATIONS CANONICAL MODEL EVALUATION

/evaluations/ling-3.0-tiny

Ling 3.0 Tiny

Published evaluation evidence and testing history. Results remain attributed to their original configurations and methods.

latest evidence 2026-09-25

Overview

inclusionAI (Ant Group) · 2 tested configurations · 4 evidence events · newest evidence 2026-09-25.

Contents: Findings · Event results · Tested configurations · Chronology · Technical provenance

Findings

Each finding is attributed to the event and tested configuration that measured it; exact values, failures and limitations are in the event results. A historical event's recorded status describes that event's own sections at its date and is not a verdict on newer sections.

  • Model classification: NOT_READY - attributed to the 2026-09-25 event, configuration llama.cpp official Q8_0 (Reasoning On, deployment sampler, 32K) · llama.cpp b10999 · RTX 5070.
  • Recommended configuration: llama.cpp official Q8_0 (Reasoning On, deployment sampler, 32K) (current): event result.
  • Latest evidence: 2026-09-25: event result
  • 2026-09-25 · llama.cpp official Q8_0 (Reasoning On, deployment sampler, 32K) · llama.cpp b10999 · RTX 5070: classification NOT_READY. Event result
  • 2026-09-25 · llama.cpp official Q8_0 (Reasoning Off, deployment sampler, 32K) · llama.cpp b10999 · RTX 5070: classification NOT_READY. Event result
  • 2026-09-25 · llama.cpp official Q8_0 (Reasoning Off, deployment sampler, 32K) · llama.cpp b10999 · RTX 5070: classification NOT_READY. Event result
  • 2026-09-24 · llama.cpp official Q8_0 (Reasoning On, deployment sampler, 32K) · llama.cpp b10999 · RTX 5070: classification NOT_READY. Event result

Event results

Newest first. Each event is one independent testing/publication event with its full report linked once at the top of its section; scores, failures and limitations stay visible and raw section-provenance JSON stays collapsed.

2026-09-25: Reasoning On classification re-derived under the 2026-09-25 reasoning-profiles snapshot from retained hash-bound evidence (no new inference)

WELP Reasoning-Profile Re-derivation: Ling 3.0 Tiny (Reasoning On, publisher default) · config llama.cpp official Q8_0 (Reasoning On, deployment sampler, 32K) · llama.cpp b10999 · RTX 5070 · maturity: CURRENT_WELP · status: NOT_READY

Configuration: llama.cpp official Q8_0 (reasoning-on, deployment sampler, 32K).

Full report (REPORT.md)

WELP outcome

  • Outcome: COMPLETE_PASS
  • Classification: not recorded

Publication state: published; canonical evidence linked above

Quality and capabilities

  • Constrained result: not recorded
Guardrails and limitations

Reliability: not recorded

LocalMaxxing

  • Status: SUBMITTED
  • Submission reference: not recorded
  • verifiedRun: null (not claimed)

the engine-native llama-bench record does not exercise the reasoning control, so both reasoning profiles share this record

Tested condition: llama.cpp b10999 · RTX 5070.

Full tested identity, runtime and context details (exact hashes, versions and configuration)

Identity

FieldValue
Profile identityling-3.0-tiny-q8-0-llamacpp-b10999-rtx5070-deployment
ModelLing 3.0 Tiny
Producernot recorded
Tested artifactnot recorded
Precisionnot recorded
Artifact SHA-2569299a9e5cbc540597619e252a41fd671faa4e84e619e3cea816542c84e19f0d6
Campaignling-3.0-tiny-rtx5070-welp-reasoning-on-2026-09-25
Record date2026-09-25

Runtime and hardware

FieldValue
Enginellama.cpp b10999
Runtime tagb10999
Runtime commitb04d4e567cd2fb8d2ded6e17d38dbbcfafe29063
Runtime buildbuild-cuda-sm120
HardwareNVIDIA GeForce RTX 5070 12 GB

Context profile

FieldValue
Practical defaultnot recorded
Guarded contextnot recorded
Native model-card maximumnot recorded
Model-card envelope completenot recorded
Native maximum dispositionnot recorded

2026-09-25: Full WELP characterization of the Reasoning Off deployment profile: independent setup/calibration, same frozen task set, dual-profile completion of the case-C model

WELP Reasoning-Profile Characterization: Ling 3.0 Tiny (Reasoning Off, supported alternate) · config llama.cpp official Q8_0 (Reasoning Off, deployment sampler, 32K) · llama.cpp b10999 · RTX 5070 · maturity: CURRENT_WELP · status: NOT_READY

Configuration: llama.cpp official Q8_0 (Reasoning Off, deployment sampler, 32K).

Full report (REPORT.md)

WELP outcome

  • Outcome: COMPLETE_PASS
  • Classification: not recorded

Publication state: published; canonical evidence linked above

Quality and capabilities

  • Constrained result: not recorded
Guardrails and limitations

Reliability: not recorded

LocalMaxxing

  • Status: SUBMITTED
  • Submission reference: not recorded
  • verifiedRun: null (not claimed)

this profile's own bench re-measurement; consistent with the existing service record (benchmark does not exercise the reasoning control; no duplicate submitted)

Tested condition: llama.cpp b10999 · RTX 5070.

Full tested identity, runtime and context details (exact hashes, versions and configuration)

Identity

FieldValue
Profile identityling-3.0-tiny-q8-0-llamacpp-b10999-rtx5070-deployment-reasoning-off
ModelLing 3.0 Tiny
Producernot recorded
Tested artifactnot recorded
Precisionnot recorded
Artifact SHA-2569299a9e5cbc540597619e252a41fd671faa4e84e619e3cea816542c84e19f0d6
Campaignling-3.0-tiny-rtx5070-welp-reasoning-off-2026-09-25
Record date2026-09-25

Runtime and hardware

FieldValue
Enginellama.cpp b10999
Runtime tagb10999
Runtime commitb04d4e567cd2fb8d2ded6e17d38dbbcfafe29063
Runtime buildbuild-cuda-sm120
HardwareNVIDIA GeForce RTX 5070 12 GB

Context profile

FieldValue
Practical defaultnot recorded
Guarded contextnot recorded
Native model-card maximumnot recorded
Model-card envelope completenot recorded
Native maximum dispositionnot recorded

2026-09-25: First WELP Agentic section execution: three bounded tool tasks through the pinned harness in a disposable offline sandbox; methodology-validation event on a known model

WELP Agentic: Ling 3.0 Tiny (Reasoning Off; methodology validation) · config llama.cpp official Q8_0 (Reasoning Off, deployment sampler, 32K) · llama.cpp b10999 · RTX 5070 · maturity: CURRENT_WELP · status: NOT_READY

Configuration: llama.cpp official Q8_0 (Reasoning Off, deployment sampler, 32K).

Full report (REPORT.md)

WELP outcome

  • Outcome: COMPLETE_PASS
  • Classification: not recorded

Publication state: published; canonical evidence linked above

Quality and capabilities

  • Constrained result: not recorded
Guardrails and limitations

Reliability: not recorded

LocalMaxxing

  • Status: SUBMITTED
  • Submission reference: not recorded
  • verifiedRun: null (not claimed)

this profile's own bench re-measurement; consistent with the existing service record (benchmark does not exercise the reasoning control; no duplicate submitted)

Tested condition: llama.cpp b10999 · RTX 5070.

Full tested identity, runtime and context details (exact hashes, versions and configuration)

Identity

FieldValue
Profile identityling-3.0-tiny-q8-0-llamacpp-b10999-rtx5070-deployment-reasoning-off
ModelLing 3.0 Tiny
Producernot recorded
Tested artifactnot recorded
Precisionnot recorded
Artifact SHA-2569299a9e5cbc540597619e252a41fd671faa4e84e619e3cea816542c84e19f0d6
Campaignling-3.0-tiny-rtx5070-welp-reasoning-off-2026-09-25
Record date2026-09-25

Runtime and hardware

FieldValue
Enginellama.cpp b10999
Runtime tagb10999
Runtime commitb04d4e567cd2fb8d2ded6e17d38dbbcfafe29063
Runtime buildbuild-cuda-sm120
HardwareNVIDIA GeForce RTX 5070 12 GB

Context profile

FieldValue
Practical defaultnot recorded
Guarded contextnot recorded
Native model-card maximumnot recorded
Model-card envelope completenot recorded
Native maximum dispositionnot recorded

2026-09-24: Second WELP stabilization-cohort campaign: hybrid-attention MoE characterized end-to-end on the pinned runtime (official Q8_0)

WELP Initial Evaluation: inclusionAI Ling 3.0 Tiny (reasoning-on) · config llama.cpp official Q8_0 (Reasoning On, deployment sampler, 32K) · llama.cpp b10999 · RTX 5070 · maturity: CURRENT_WELP · status: NOT_READY

Configuration: llama.cpp official Q8_0 (reasoning-on, deployment sampler, 32K).

Full report (REPORT.md)

WELP outcome

  • Outcome: COMPLETE_PASS
  • Classification: not recorded

Publication state: published; canonical evidence linked above

Quality and capabilities

  • Constrained result: not recorded
Guardrails and limitations

Reliability: not recorded

LocalMaxxing

  • Status: SUBMITTED
  • Submission reference: not recorded
  • verifiedRun: null (not claimed)

Tested condition: llama.cpp b10999 · RTX 5070.

Full tested identity, runtime and context details (exact hashes, versions and configuration)

Identity

FieldValue
Profile identityling-3.0-tiny-q8-0-llamacpp-b10999-rtx5070-deployment
ModelLing 3.0 Tiny
Producernot recorded
Tested artifactnot recorded
Precisionnot recorded
Artifact SHA-2569299a9e5cbc540597619e252a41fd671faa4e84e619e3cea816542c84e19f0d6
Campaignling-3.0-tiny-rtx5070-welp-characterization-2026-09-24
Record date2026-09-24

Runtime and hardware

FieldValue
Enginellama.cpp b10999
Runtime tagb10999
Runtime commitb04d4e567cd2fb8d2ded6e17d38dbbcfafe29063
Runtime buildbuild-cuda-sm120
HardwareNVIDIA GeForce RTX 5070 12 GB

Context profile

FieldValue
Practical defaultnot recorded
Guarded contextnot recorded
Native model-card maximumnot recorded
Model-card envelope completenot recorded
Native maximum dispositionnot recorded

Tested configurations

Exact profile identities, sampling conditions and provenance are in each configuration's details below.

ConfigurationArtifact / precisionRuntimeHardwareContext / reasoningEvents
llama.cpp official Q8_0 (Reasoning On, deployment sampler, 32K) · llama.cpp b10999 · RTX 5070Q8_0 official publisher GGUFllama.cpp b10999RTX 5070not recorded2 (2026-09-24, 2026-09-25)
llama.cpp official Q8_0 (Reasoning Off, deployment sampler, 32K) · llama.cpp b10999 · RTX 5070Q8_0 official publisher GGUFllama.cpp b10999RTX 5070not recorded2 (2026-09-25, 2026-09-25)
Configuration details: llama.cpp official Q8_0 (Reasoning On, deployment sampler, 32K) · llama.cpp b10999 · RTX 5070
  • Full profile name: llama.cpp official Q8_0 (Reasoning On, deployment sampler, 32K).
  • Profile identity: ling-3.0-tiny-q8-0-llamacpp-b10999-rtx5070-deployment.
  • Status: current canonical/recommended tested surface.
  • Deployment topology (as recorded): Full-GPU upstream llama.cpp CUDA serving on RTX 5070 12GB: -ngl 99 (24/24 layers), -fa on, f16 KV, -np 1, embedded GGUF Bailing-V3 Jinja template, DEPLOYMENT sampler (temp 0.2 / top_k 80 / top_p 1.0 / repeat_penalty 1.05); effective reasoning ON (bailing v3 thinking default; enable_thinking=false OFF control proven effective but not the campaign profile). Default context 32,768 / native exact 131,072 / advertised 262,144 (YaRN x2.0, mechanism-qualified on this runtime); uncached scientific arms via --cache-ram 0 --no-cache-prompt --slot-prompt-similarity 0 with cache_prompt:false (proven by repeated-prompt probe). Publisher sampler (temp 1.0 / top_k 20 / top_p 0.95) measured as a matched comparison, never pooled.
  • Artifact (as recorded): inclusionAI/Ling-3.0-tiny-GGUF@01b850e2 :: Ling-3.0-tiny-Q8_0.gguf (sha256 9299a9e5cbc540597619e252a41fd671faa4e84e619e3cea816542c84e19f0d6)
  • Runtime (as recorded): llama.cpp b10999.
  • Hardware (as recorded): NVIDIA GeForce RTX 5070 12 GB.
  • Profile Metadata

Events on this configuration:

Configuration details: llama.cpp official Q8_0 (Reasoning Off, deployment sampler, 32K) · llama.cpp b10999 · RTX 5070
  • Full profile name: llama.cpp official Q8_0 (Reasoning Off, deployment sampler, 32K).
  • Profile identity: ling-3.0-tiny-q8-0-llamacpp-b10999-rtx5070-deployment-reasoning-off.
  • Status: validated alternate tested surface.
  • Artifact (as recorded): inclusionAI/Ling-3.0-tiny-GGUF@01b850e2 :: Ling-3.0-tiny-Q8_0.gguf (sha256 9299a9e5cbc540597619e252a41fd671faa4e84e619e3cea816542c84e19f0d6)
  • Runtime (as recorded): llama.cpp b10999.
  • Hardware (as recorded): NVIDIA GeForce RTX 5070 12 GB.
  • Profile Metadata

Events on this configuration:

Chronology

Compact index; full results are in Event results.

DateEventConfigurationStatus
2026-09-25Reasoning On classification re-derived under the 2026-09-25 reasoning-profiles snapshot from retained hash-bound evidence (no new inference)llama.cpp official Q8_0 (Reasoning On, deployment sampler, 32K) · llama.cpp b10999 · RTX 5070NOT_READY
2026-09-25Full WELP characterization of the Reasoning Off deployment profile: independent setup/calibration, same frozen task set, dual-profile completion of the case-C modelllama.cpp official Q8_0 (Reasoning Off, deployment sampler, 32K) · llama.cpp b10999 · RTX 5070NOT_READY
2026-09-25First WELP Agentic section execution: three bounded tool tasks through the pinned harness in a disposable offline sandbox; methodology-validation event on a known modelllama.cpp official Q8_0 (Reasoning Off, deployment sampler, 32K) · llama.cpp b10999 · RTX 5070NOT_READY
2026-09-24Second WELP stabilization-cohort campaign: hybrid-attention MoE characterized end-to-end on the pinned runtime (official Q8_0)llama.cpp official Q8_0 (Reasoning On, deployment sampler, 32K) · llama.cpp b10999 · RTX 5070NOT_READY

Technical provenance

Publication, acceptance and source-pin provenance for each event, collapsed; the scientific content above never depends on it.

Model evidence tree (GitHub)

Event provenance: 2026-09-25 WELP Reasoning-Profile Re-derivation: Ling 3.0 Tiny (Reasoning On, publisher default)
  • Campaign: ling-3.0-tiny-rtx5070-welp-reasoning-on-2026-09-25
  • Export disposition: WEBSITE_PUBLISHED; publication date 2026-09-25
  • Canonical citation (immutable commit/path pin): https://github.com/WumboLabs/evaluations/blob/a5717d2fecc21184da6dfa20d973fb8e4fbca0e4/models/ling-3.0-tiny/events/ling-3-0-tiny-rtx5070-welp-reasoning-on-20260925/REPORT.md
  • Export identity: event_date 2026-09-25, event_id ling-3-0-tiny-rtx5070-welp-reasoning-on-20260925, event_type follow-up, evidence_scope ['reliability', 'context', 'capabilities', 'performance', 'localmaxxing', 'model-classification'], model_id ling-3.0-tiny, profile_id ling-3.0-tiny-q8-0-llamacpp-b10999-rtx5070-deployment, profile_status current, related_profile_ids ['ling-3-0-tiny-q8-0-llamacpp-b10999-rtx5070-deployment-reasoning-off']
  • Export source pin: repository-cache WumboLabs/evaluations @ c41e09b5909f633e3d9093747398e162e17a50c4 path models/ling-3.0-tiny/events/ling-3-0-tiny-rtx5070-welp-reasoning-on-20260925/website-publication.json sha256 84a8a7153682ff4d311ffc082b04758ddae1662a73983164c309b14dfb74f2e3

Publication provenance note (as recorded): Reuse/re-derivation event under welp-reasoning-topology-0.1.0-draft: full migration matrix (REUSE/RE-DERIVED/INCOMPARABLE) in the campaign's REUSE-MIGRATION.md; the source event stays immutable. Executed under snapshot welp-next-snapshot-2026-09-25-reasoning-profiles (WELP commit 193cb645c81440cc1232447f3ebc32a8197a7f3e, DRAFT - NOT v1.0).

Event provenance: 2026-09-25 WELP Reasoning-Profile Characterization: Ling 3.0 Tiny (Reasoning Off, supported alternate)
  • Campaign: ling-3.0-tiny-rtx5070-welp-reasoning-off-2026-09-25
  • Export disposition: WEBSITE_PUBLISHED; publication date 2026-09-25
  • Canonical citation (immutable commit/path pin): https://github.com/WumboLabs/evaluations/blob/a5717d2fecc21184da6dfa20d973fb8e4fbca0e4/models/ling-3.0-tiny/events/ling-3-0-tiny-rtx5070-welp-reasoning-off-20260925/REPORT.md
  • Export identity: event_date 2026-09-25, event_id ling-3-0-tiny-rtx5070-welp-reasoning-off-20260925, event_type follow-up, evidence_scope ['reliability', 'context', 'capabilities', 'performance', 'localmaxxing', 'model-classification'], model_id ling-3.0-tiny, profile_id ling-3.0-tiny-q8-0-llamacpp-b10999-rtx5070-deployment-reasoning-off, profile_status current-alternate, related_profile_ids ['ling-3-0-tiny-q8-0-llamacpp-b10999-rtx5070-deployment']
  • Export source pin: repository-cache WumboLabs/evaluations @ c41e09b5909f633e3d9093747398e162e17a50c4 path models/ling-3.0-tiny/events/ling-3-0-tiny-rtx5070-welp-reasoning-off-20260925/website-publication.json sha256 5a5374c9e13a9ca7fcc89e7e8bbf48be18d7d5933f42685988b1f309192959a3

Publication provenance note (as recorded): Full independent WELP profile campaign under welp-next-snapshot-2026-09-25-reasoning-profiles (WELP commit 193cb645c81440cc1232447f3ebc32a8197a7f3e, DRAFT - NOT v1.0); artifact restored from the verified canonical archive with SHA-256 re-verification; LocalMaxxing VERIFIED_EXISTING (the llama-bench record does not exercise the reasoning control).

Event provenance: 2026-09-25 WELP Agentic: Ling 3.0 Tiny (Reasoning Off; methodology validation)
  • Campaign: ling-3.0-tiny-rtx5070-welp-reasoning-off-2026-09-25
  • Export disposition: WEBSITE_PUBLISHED; publication date 2026-09-25
  • Canonical citation (immutable commit/path pin): https://github.com/WumboLabs/evaluations/blob/599063be57edd4f3b01f27c33da02d48abc541d8/models/ling-3.0-tiny/events/ling-3-0-tiny-rtx5070-welp-agentic-20260925/REPORT.md
  • Export identity: event_date 2026-09-25, event_id ling-3-0-tiny-rtx5070-welp-agentic-20260925, event_type welp-agentic-methodology-validation, evidence_scope ['agentic-task-outcomes', 'agentic-methodology-validation'], model_id ling-3.0-tiny, profile_id ling-3.0-tiny-q8-0-llamacpp-b10999-rtx5070-deployment-reasoning-off, profile_status current-alternate
  • Export source pin: repository-cache WumboLabs/evaluations @ 0382fc7dd89fd403869fa3d81830adc15233da7b path models/ling-3.0-tiny/events/ling-3-0-tiny-rtx5070-welp-agentic-20260925/website-publication.json sha256 5044e15ea13bbfc65819cac929429e7357b2ce43f8bff52dba2ee8f32f049ce6

Publication provenance note (as recorded): METHODOLOGY_VALIDATION event under welp-next-snapshot-2026-09-25-model-agentic (WELP commit 5f611de36e44a2d06fa41acd75aa62dcf66c0edb, DRAFT - NOT v1.0). Agentic verdict reported independently of the Model verdict; never averaged.

Event provenance: 2026-09-24 WELP Initial Evaluation: inclusionAI Ling 3.0 Tiny (reasoning-on)
  • Campaign: ling-3.0-tiny-rtx5070-welp-characterization-2026-09-24
  • Export disposition: WEBSITE_PUBLISHED; publication date 2026-09-24
  • Canonical citation (immutable commit/path pin): https://github.com/WumboLabs/evaluations/blob/43f56d8add4b3dfc51183ccd818fc6037e79bd9c/models/ling-3.0-tiny/events/ling-3-0-tiny-rtx5070-welp-characterization-20260924/REPORT.md
  • Export identity: event_date 2026-09-24, event_id ling-3-0-tiny-rtx5070-welp-characterization-20260924, event_type initial-evaluation, model_id ling-3.0-tiny, profile_id ling-3.0-tiny-q8-0-llamacpp-b10999-rtx5070-deployment, profile_status current
  • Export source pin: repository-cache WumboLabs/evaluations @ 719810861ab6d02136ea68fdd5d5e4e34e3e4ddb path models/ling-3.0-tiny/events/ling-3-0-tiny-rtx5070-welp-characterization-20260924/website-publication.json sha256 128e900f0a2f3965b0833ad30c7ec5f5654a82b07e3a3cea9571ba2ca0f83871

Publication provenance note (as recorded): Native current-WELP event under the consolidated architecture; the campaign completed on the success branch with zero methodology defects (WELP stabilization campaign CLEAN, clean-count 2). Executed under snapshot welp-next-snapshot-2026-09-24-context-outcome-repair (WELP commit ead28bde1ea9bce20cec6b3930125016068a8bbc, DRAFT - NOT v1.0); classification contract welp-final-classification-0.4.0-draft; verdict NOT_READY (R-C3) with campaign execution COMPLETE_PASS.