---
title: "LFM2.5 text family | Ethical Stack"
url: https://ethicalstack.ai/models/lfm2-5/
catalog_version: f6baff2119d0e9158520da7fbe4232011dac79f25bd08757436d7a1dbd593db2
---

# LFM2.5 text family

Small text models for local chat, reasoning and tool use, from 230M to 24B total parameters. Includes Japanese and mixture-of-experts variants.

Canonical profile: https://ethicalstack.ai/models/lfm2-5/

Scope: Current LFM2.5 text models: 230M, 350M, 1.2B Instruct/Thinking/JP, 8B-A1B and 24B-A2B; superseded 2.6B is excluded.

Published overall: 4.3; 6/6 dimensions rated.

Round the equal-weight mean of available numerical ratings to one decimal; require at least four of six for an overall score. BDS status stays outside the mean; for eligible profiles, a boycott target caps the final score at 2/10 and a pressure target subtracts one point (minimum 1). Not listed and not checked do not adjust the score.

## Energy practices

Finding: 4/10

Assessment: https://ethicalstack.ai/models/lfm2-5/#energy

Local deployment and smaller models give you useful controls; clean-electricity evidence is missing.

Reason: 1 + 1 reporting + 0 electricity + 2 controls = 4/10. This grades documented practices and transparency.

Argument: This rating credits documented practices, not assumed energy savings. The model library and local guides document hardware/runtime choices, smaller checkpoints and quantized formats where supported. Actual consumption depends on the chosen model, device and workload.

Type: argument; confidence: medium

Scope (dimension): Current LFM2.5 text models: 230M, 350M, 1.2B Instruct/Thinking/JP, 8B-A1B and 24B-A2B; superseded 2.6B is excluded.

Research reviewed: not recorded; evidence date: not recorded

Unknowns: Comparable task-level electricity, water, training footprint and independently verified clean-power sourcing are not published for this choice.

```json
{
  "method": "practices-transparency-v1",
  "scope": "Current LFM2.5 text models: 230M, 350M, 1.2B Instruct/Thinking/JP, 8B-A1B and 24B-A2B; superseded 2.6B is excluded.",
  "components": {
    "disclosure": {
      "score": 1,
      "reason": "The documented route identifies where inference can run, without an audited environmental footprint.",
      "sourceUrls": [
        "https://huggingface.co/LiquidAI/LFM2.5-1.2B-Instruct"
      ]
    },
    "electricity": {
      "score": 0,
      "reason": "No attributable electricity procurement or hourly clean-power evidence was found in the cited material.",
      "sourceUrls": [
        "https://huggingface.co/LiquidAI/LFM2.5-1.2B-Instruct",
        "https://www.liquid.ai/company"
      ]
    },
    "demand": {
      "score": 2,
      "reason": "Model/runtime and size choices provide several controls, without measured energy savings.",
      "sourceUrls": [
        "https://docs.liquid.ai/lfm/models/complete-library",
        "https://docs.liquid.ai/deployment/on-device/llama-cpp"
      ]
    }
  },
  "actualImpactMeasured": false
}
```

- [LFM2.5 text family official release documentation](https://huggingface.co/LiquidAI/LFM2.5-1.2B-Instruct) (primary; checked 2026-10-28)

- [Liquid AI complete model library](https://docs.liquid.ai/lfm/models/complete-library) (primary; checked 2026-10-28)

- [Liquid AI local inference guide](https://docs.liquid.ai/deployment/on-device/llama-cpp) (primary; checked 2026-10-28)

- [Liquid AI API, license and custom deployment pricing](https://www.liquid.ai/pricing) (primary; checked 2026-10-28)

- [Liquid AI company values](https://www.liquid.ai/company) (primary; checked 2026-10-28)

## Training data

Finding: 4/10

Assessment: https://ethicalstack.ai/models/lfm2-5/#training-rights

Liquid describes broad training sources, but has not published a permission audit for the whole training mix.

Reason: 1 + 1 source transparency + 1 permission + 1 creator controls = 4/10, with the permission cap applied.

Argument: The model documentation describes its base model and training approach. Liquid’s privacy policy adds public internet material and business datasets by agreement. That gives some provenance and a limited permission basis, but no itemized evidence clearing all underlying text, images, speech or inherited training data. The model license grants use of weights; it does not settle training-data rights.

Type: argument; confidence: medium

Scope (profile): Current LFM2.5 text models: 230M, 350M, 1.2B Instruct/Thinking/JP, 8B-A1B and 24B-A2B; superseded 2.6B is excluded.

Research reviewed: not recorded; evidence date: not recorded

Unknowns: Dataset-level licenses, consent for speech/images, rights checks for inherited training data, and removal outcomes remain unclear.

```json
{
  "method": "training-rights-evidence-v1",
  "scope": "The stated training release or selected model",
  "components": {
    "provenance": {
      "score": 1,
      "reason": "Broad source classes and model lineage are described; no complete dataset ledger is published.",
      "sourceUrls": [
        "https://huggingface.co/LiquidAI/LFM2.5-1.2B-Instruct",
        "https://www.liquid.ai/privacy-policy"
      ]
    },
    "permissions": {
      "score": 1,
      "reason": "Business dataset agreements provide a limited stated permission basis; the wider public-web mix has no inspectable permission audit.",
      "sourceUrls": [
        "https://www.liquid.ai/privacy-policy"
      ]
    },
    "creatorControls": {
      "score": 1,
      "reason": "The privacy policy offers personal-data objections, including training use where applicable; this is narrower than a general creator opt-out.",
      "sourceUrls": [
        "https://www.liquid.ai/privacy-policy"
      ]
    }
  },
  "routeDependency": null,
  "modelReference": null
}
```

- [LFM2.5 text family official release documentation](https://huggingface.co/LiquidAI/LFM2.5-1.2B-Instruct) (primary; checked 2026-10-28)

- [Liquid AI privacy policy](https://www.liquid.ai/privacy-policy) (primary; checked 2026-10-28)

- [LFM Open License guide](https://docs.liquid.ai/lfm/help/model-license) (primary; checked 2026-10-28)

## Your data & privacy

Finding: 8/10

Assessment: https://ethicalstack.ai/models/lfm2-5/#privacy

Running locally can keep prompts on your own hardware; your app and tools still control other data paths.

Reason: 8/10 credits a documented local data path with configuration and audit limits.

Argument: Downloadable weights can be run without sending prompts to Liquid once setup is complete. The runtime, remote tools, logs and backups remain your responsibility.

Type: argument; confidence: medium

Scope (profile): Current LFM2.5 text models: 230M, 350M, 1.2B Instruct/Thinking/JP, 8B-A1B and 24B-A2B; superseded 2.6B is excluded.

Research reviewed: not recorded; evidence date: not recorded

Unknowns: No independent network audit is claimed. App backups, runtime telemetry, tools and hosting setup can create other copies.

- [LFM2.5 text family official release documentation](https://huggingface.co/LiquidAI/LFM2.5-1.2B-Instruct) (primary; checked 2026-10-28)

- [Liquid AI privacy policy](https://www.liquid.ai/privacy-policy) (primary; checked 2026-10-28)

- [Liquid AI local inference guide](https://docs.liquid.ai/deployment/on-device/llama-cpp) (primary; checked 2026-10-28)

- [Liquid Console and API terms](https://www.liquid.ai/terms-conditions) (primary; checked 2026-10-28)

## Openness

Finding: 6/10

Assessment: https://ethicalstack.ai/models/lfm2-5/#openness

Weights are downloadable, with commercial use restricted by annual revenue.

Reason: 6/10 credits inspectable downloadable weights and practical deployment, with a commercial revenue limit and incomplete training disclosure.

Argument: Liquid provides weights, model cards and runtime instructions, but uses a custom license. Commercial users at $10M annual revenue or above need a separate license, so these weights do not offer unrestricted Apache-style commercial use. The complete training data is also unavailable.

Type: argument; confidence: medium

Scope (profile): Current LFM2.5 text models: 230M, 350M, 1.2B Instruct/Thinking/JP, 8B-A1B and 24B-A2B; superseded 2.6B is excluded.

Research reviewed: not recorded; evidence date: not recorded

Unknowns: The full training dataset and reproducible training pipeline are unavailable.

- [LFM2.5 text family official release documentation](https://huggingface.co/LiquidAI/LFM2.5-1.2B-Instruct) (primary; checked 2026-10-28)

- [LFM Open License guide](https://docs.liquid.ai/lfm/help/model-license) (primary; checked 2026-10-28)

- [Liquid Finetune official repository](https://github.com/Liquid4All/liquid-finetune) (primary; checked 2026-10-28)

- [Pipette official benchmark clients](https://github.com/Liquid4All/pipette-clients) (primary; checked 2026-10-28)

## Worker protections

Finding: 2/10

Assessment: https://ethicalstack.ai/models/lfm2-5/#labour

Liquid publishes hiring and wellbeing commitments, but little evidence about its wider AI workforce.

Reason: 1 + 1 + 0 + 0 = 2/10. Documented worker protections; missing disclosure is an evidence gap.

Argument: Its careers page commits to equal opportunity and interview accommodations; company values mention wellbeing. These earn limited policy credit. The reviewed sources provide no contractor pay standards, protected grievance channel or independent evidence covering annotation, moderation and other AI supply-chain work. This entry carries the same developer/operator finding to LFM2.5 text family; it does not establish that this product was deployed under every cited relationship.

Type: argument; confidence: medium

Scope (dimension): Liquid AI: the documented developer/operator record, not proof of deployment of LFM2.5 text family.

Research reviewed: not recorded; evidence date: not recorded

Unknowns: Contractor conditions, worker representation, grievance outcomes and independent workforce audits remain undisclosed.

```json
{
  "method": "worker-safeguards-evidence-v1",
  "entity": "liquid-ai",
  "company": "Liquid AI",
  "scope": "Liquid AI employees and AI supply-chain workers",
  "components": {
    "conditions": {
      "score": 1,
      "reason": "Public equal-opportunity and accommodation commitments provide basic policy evidence.",
      "sourceUrls": [
        "https://www.liquid.ai/careers"
      ]
    },
    "voice": {
      "score": 0,
      "reason": "No worker grievance or representation safeguard was found in the reviewed public material.",
      "sourceUrls": [
        "https://www.liquid.ai/company",
        "https://www.liquid.ai/careers"
      ]
    },
    "assurance": {
      "score": 0,
      "reason": "No independent worker-level implementation review was supplied in these sources.",
      "sourceUrls": [
        "https://www.liquid.ai/company",
        "https://www.liquid.ai/careers"
      ]
    }
  },
  "disclosureLimited": true
}
```

- [Liquid AI careers and equal-opportunity statement](https://www.liquid.ai/careers) (primary; checked 2026-10-28)

- [Liquid AI company values](https://www.liquid.ai/company) (primary; checked 2026-10-28)

## Military ties

Finding: 2/10

Assessment: https://ethicalstack.ai/models/lfm2-5/#military-contracts

Liquid actively offers tactical AI for threat detection, ISR and uncrewed systems.

Reason: 1 + 1 use restrictions + 1 enforcement & oversight + 1 relationship disclosure = 4, capped at 2/10. Known defense ties lower this rating; safeguards can improve it within the relationship’s cap.

Argument: Liquid’s defense page markets on-device models for combat-relevant intelligence and autonomous systems. We treat that operational offering as direct military enablement under the rubric, which caps the score at 2. Its general service terms prohibit unlawful harm and permit suspension, but do not rule out defense use. This is evidence of an offering, not proof that any particular checkpoint has been deployed by a named military customer. This entry carries the same developer/operator finding to LFM2.5 text family; it does not establish that this product was deployed under every cited relationship.

Type: argument; confidence: medium

Scope (dimension): Liquid AI: the documented developer/operator record, not proof of deployment of LFM2.5 text family.

Research reviewed: not recorded; evidence date: not recorded

Unknowns: Named defense contracts, actual deployments, customer screening and independent oversight are not disclosed on the reviewed page.

```json
{
  "method": "military-ties-priority-v2",
  "entity": "liquid-ai",
  "company": "Liquid AI",
  "scope": "Liquid AI’s advertised defense offering; no specific model deployment inferred",
  "components": {
    "restrictions": {
      "score": 1,
      "reason": "General prohibitions cover unlawful harm; no blanket restriction on operational military use is stated.",
      "sourceUrls": [
        "https://www.liquid.ai/terms-conditions"
      ]
    },
    "oversight": {
      "score": 1,
      "reason": "Terms allow suspension for violations, without a disclosed military-use review process.",
      "sourceUrls": [
        "https://www.liquid.ai/terms-conditions"
      ]
    },
    "transparency": {
      "score": 1,
      "reason": "The defense offering is public, but customers, contract scope and deployments are not named.",
      "sourceUrls": [
        "https://www.liquid.ai/defense"
      ]
    }
  },
  "disclosureLimited": false,
  "relationship": "direct-operational",
  "relationshipReason": "The vendor explicitly offers tactical threat detection, ISR, dismounted-team decision support and onboard perception for autonomous uncrewed systems; this is operational military enablement, without a claim of a named contract.",
  "score": 2,
  "baseScore": 1,
  "safeguardPoints": 3,
  "safeguardAdjustment": 3,
  "rawScore": 4,
  "cap": 2
}
```

- [Liquid AI defense offering](https://www.liquid.ai/defense) (primary; checked 2026-10-28)

- [Liquid Console and API terms](https://www.liquid.ai/terms-conditions) (primary; checked 2026-10-28)

## BDS targets

Finding: Not listed

Assessment: https://ethicalstack.ai/models/lfm2-5/#israel-palestine

Liquid AI is not named in the checked BNC boycott and technology target lists.

Reason: Campaign category, excluded from numerical scoring.

Argument: The BNC consumer guide and No Tech campaign page do not name Liquid AI as a boycott or pressure target. This is a limited campaign-list finding; it is not an endorsement or a wider assessment of every supplier or customer. The defense offering is assessed separately in military ties. This entry carries the same developer/operator finding to LFM2.5 text family; it does not establish that this product was deployed under every cited relationship.

Type: argument; confidence: medium

Scope (dimension): Liquid AI: the documented developer/operator record, not proof of deployment of LFM2.5 text family.

Research reviewed: not recorded; evidence date: not recorded

Unknowns: The checked lists do not identify every commercial relationship or local campaign.

- [BNC boycott and pressure guide](https://bdsmovement.net/Guide-to-BDS-Boycott) (primary; checked 2026-10-28)

- [BNC No Tech campaign targets](https://bdsmovement.net/no-tech-oppression-apartheid-or-genocide) (primary; checked 2026-10-28)

## Reading the evidence

- Assessments are provisional editorial judgments. Cite the profile and the underlying sources, and retain confidence, scope and unknowns.

- Unknown, not-applicable and model-dependent are different states. Missing evidence does not establish good or bad conduct.

- Company-level evidence does not establish use of a particular product or model. Account plan, model, host and route can change an assessment.

- Energy scores describe documented practices and transparency, not measured request emissions. Zero component points can mean an evidence gap.

- Published overall scores require at least four of six numerical dimensions. Preserve coverage and BDS adjustments; BDS itself is categorical.

- No named BDS target found is limited to the reviewed campaign sources and is not approval. Personalized priorities do not change the published overall score.

- Review and source-check dates describe the research, not this deployment. Fetch current primary sources when the decision requires current verification.

Methodology: https://ethicalstack.ai/methodology/