Best Named Entity Recognition Models to Run Locally in 2026: LFM2.5-Encoder-350M-PII-Detector

Quick Answer

LiquidAI’s LFM2.5-Encoder-350M-PII-Detector is the top pick as of September 2026 because it is the newest eligible release; the selection includes older models because too few releases fall within the past year [1][2][4][5][6]. In order, the ranking is LFM2.5-Encoder-350M-PII-Detector (LiquidAI), privacy-filter (OpenAI), gliner-PII (NVIDIA), llmlingua-2-bert-base-multilingual-cased-meetingbank (Microsoft), and llmlingua-2-xlm-roberta-large-meetingbank (Microsoft) [1][2].

Key Takeaways

• Ranking note: No public benchmark scores any candidate, so the order is newest release first, then downloads.[1][2][4][5][6] Too few releases fall within the past 12 months, so the newest available family entries are included.[2][4][5] Scope: this ranking admits only token classification models from labs with a published paper or leaderboard record, using the Hugging Face token-classification pipeline tag.[1][2][3][4][5][6]
• LiquidAI’s LFM2.5-Encoder-350M-PII-Detector ranks first on release recency, with publication on 2026-06-12.[6] Parameters: 355M; 4-bit weight memory: 177.5 MB, estimated (params x 0.5 bytes).[6] Context: 128K tokens; license: lfm1.0; no public benchmark yet.[6]
• OpenAI’s privacy-filter ranks second, published on 2026-04-17.[1] Parameters: 1.4B; 4-bit weight memory: 0.7 GB, estimated (params x 0.5 bytes).[1] Context: 128K tokens; license: Apache-2.0; no public benchmark yet.[1]
• NVIDIA’s gliner-PII ranks third, published on 2025-08-29 under the nvidia-open-model-license.[5] Parameter count and context length are unspecified here, so a weight-memory estimate is omitted; no public benchmark yet.[5]
• Microsoft’s llmlingua-2-bert-base-multilingual-cased-meetingbank ranks fourth: its 2024-03-17 publication ties the other Microsoft entry, while downloads break the tie—292,286 versus 185,818 over the last 30 days as of 2026-09-25.[2][4] Parameters: 177M; 4-bit weight memory: 88.5 MB, estimated (params x 0.5 bytes).[2] Context: 512 tokens; license: Apache-2.0; no public benchmark yet.[2]
• Microsoft’s llmlingua-2-xlm-roberta-large-meetingbank ranks fifth, published on 2024-03-17.[4] Parameters: 559M; 4-bit weight memory: 279.5 MB, estimated (params x 0.5 bytes).[4] Context: 514 tokens; license: MIT; no public benchmark yet.[4] The accompanying paper describes prompt compression, which matters when assessing suitability for named entity recognition.[3]

How do local named entity recognition models compare on specifications and public benchmarks?

Which named entity recognition models can you run locally?

1. LFM2.5-Encoder-350M-PII-Detector

LFM2.5-Encoder-350M-PII-Detector by LiquidAI ranks first because its Hugging Face publication date, June 12, 2026, is the newest among the eligible candidates.[6][1][2][4][5] The model has 355M parameters, a 128K-token context window, and F32 weights totaling 1.4 GB.[6] Its license is lfm1.0.[6] Consider it for local personally identifiable information (PII) detection when a long input window matters; its position reflects release recency, with no public benchmark yet establishing comparative detection quality.[6]

For local hardware planning, the 355M parameters imply approximately 177.5 MB of 4-bit weight storage—estimated (params x 0.5 bytes).[6] Treat that calculation as a weight-storage estimate, not a complete GPU-memory or system-RAM requirement. A specific hardware configuration cannot be recommended from the listed specifications alone. Before deployment, verify that your chosen runtime supports the model and your intended precision, then measure memory consumption with representative inputs. The compact weight estimate should not substitute for that check.

Ranking note: the ordering is newest release first, then downloads; no public benchmark yet scores these candidates against one another. The ranking admits only token classification models from labs with a published paper or leaderboard record, using the Hugging Face token-classification pipeline tag. The family has too few releases within the last twelve months, so the newest available candidates are included, including older releases.[1][2][4][5][6]

2. privacy-filter

privacy-filter by OpenAI ranks second under the ordering rule of newest release first, then downloads.[1][6] Its Hugging Face publication date is April 17, 2026, behind LiquidAI’s LFM2.5-Encoder-350M-PII-Detector, published June 12, 2026.[1][6] No public benchmark yet establishes its relative entity-recognition accuracy, so its position reflects release timing rather than demonstrated accuracy. Consider it for local privacy-filtering workflows involving long documents, where its stated context window may be useful.[1]

OpenAI lists 1.4 billion parameters, a context length of 128K tokens, BF16 weights totaling 5.6 GB, and an Apache-2.0 license.[1] For hardware planning, 4-bit weight storage is approximately 700 MB—estimated (params x 0.5 bytes), using the cited 1.4 billion parameters.[1] Treat that calculation as a weight-storage estimate, not a complete RAM or VRAM requirement. Actual local deployment also needs memory for execution; the supplied specifications do not establish a supported quantization path or a particular GPU requirement.

The practical caveat is validation: a long context specification does not establish detection quality across document types or entity categories. Before adopting privacy-filter, evaluate representative inputs and inspect missed entities and incorrect detections. Choose hardware after measuring the complete workload’s memory use with your intended runtime and input lengths.

3. gliner-PII

gliner-PII by NVIDIA ranks third under the ordering rule “newest release first, then downloads,” behind the newer entries from LiquidAI and OpenAI.[5][6][1] No public benchmark yet establishes its position on detection quality.[5] The ranking admits only token classification models from labs with a published paper or leaderboard record. The family has too few recent releases, so the newest available are included; NVIDIA’s entry was first published on Hugging Face on August 29, 2025.[5]

The model uses the nvidia-open-model-license and recorded 23,327 downloads over the last 30 days as of September 25, 2026.[5] Download activity does not establish accuracy. Parameter count, context length and weight size are unspecified in the available specifications, leaving no defensible basis for a numerical memory estimate.[5] Check the license terms against your intended deployment before adopting the model.

For local use, treat gliner-PII as a candidate to evaluate for personally identifiable information detection. Test your entity labels and representative documents before selecting it. Hardware requirements remain unverified: a GPU recommendation or RAM minimum would be guesswork without the missing specifications and runtime measurements. Measure memory consumption in your intended runtime before provisioning hardware. The key caveat is the absence of a public benchmark: the ranking position reflects publication timing, not demonstrated superiority.[5]

4. llmlingua-2-bert-base-multilingual-cased-meetingbank

llmlingua-2-bert-base-multilingual-cased-meetingbank by Microsoft ranks fourth under the ordering rule: newest release first, then downloads.[1][2][4][5][6] Its Hugging Face publication date is March 17, 2024, placing it outside the intended recent-release window; the shortlist includes older releases because too few recent candidates qualify.[2] The release date matches Microsoft’s llmlingua-2-xlm-roberta-large-meetingbank, but its 292,286 downloads exceed that model’s 185,818 over the preceding 30 days as of September 25, 2026.[2][4] Downloads resolve that tie without establishing better extraction quality.

The model has 177M parameters, a context length of 512 tokens, and an Apache-2.0 license.[2] Published F32 weights occupy 0.7 GB.[2] For hardware planning, 4-bit weight storage is approximately 88.5 MB, estimated (params x 0.5 bytes) from the cited 177M parameters.[2] Use that estimate as a starting point for local memory budgeting, not a complete RAM or VRAM requirement or confirmation that a particular quantized runtime supports the model.

The practical use case is task-agnostic prompt compression, the task described in Microsoft’s “LLMLingua-2: Data Distillation for Efficient and Faithful Task-Agnostic Prompt Compression” paper.[3] The caveat for entity-extraction deployments is straightforward: no public benchmark yet establishes its NER standing among these candidates. Evaluate it against your entity labels before choosing it for that role.

5. llmlingua-2-xlm-roberta-large-meetingbank

llmlingua-2-xlm-roberta-large-meetingbank by Microsoft ranks fifth under the ordering rule: newest release first, then downloads. Its publication date, March 17, 2024 [4], matches Microsoft’s llmlingua-2-bert-base-multilingual-cased-meetingbank [2], but its 185,818 downloads trail that model’s 292,286 over the preceding 30 days as of September 25, 2026 [2][4]. The ranking admits only token-classification models from labs with a published paper or leaderboard record. Too few candidates meet the recent-release window, so the newest available releases from the selected families remain included [1][2][4][5][6].

The checkpoint has 559M parameters [4], a context length of 514 tokens [4], and F32 weights totaling 2.2 GB [4]. Its license is MIT [4]. For those 559M parameters [4], 4-bit weight memory is 279.5 MB, estimated (params x 0.5 bytes) [4]. Treat that calculation as a weight-storage estimate when planning local hardware; it does not establish a complete RAM or VRAM requirement, or confirm that a particular GPU can run the model.

The practical use is task-agnostic prompt compression, the purpose described in Microsoft’s LLMLingua-2: Data Distillation for Efficient and Faithful Task-Agnostic Prompt Compression [3]. The caveat for an entity-recognition shortlist is that a token-classification tag alone does not establish entity-extraction quality. Benchmark status: no public benchmark yet. Its position therefore reflects release timing and download counts, without a demonstrated named entity recognition accuracy advantage.

How much weight memory do local named entity recognition models need?

Local named entity recognition models in this selection have published weight sizes ranging from 0.7 GB to 5.6 GB, depending on the model and stored precision.[1][2][4][6] Lower-precision weight estimates should be considered separately from those published sizes.

Microsoft’s llmlingua-2-bert-base-multilingual-cased-meetingbank (Microsoft) has 177M parameters and 0.7 GB of F32 weights.[2] Its 4-bit weight memory is 88.5 MB, estimated (params x 0.5 bytes) from that parameter count.[2]

LiquidAI’s LFM2.5-Encoder-350M-PII-Detector (LiquidAI) has 355M parameters and 1.4 GB of F32 weights.[6] Its 4-bit weight memory is 177.5 MB, estimated (params x 0.5 bytes) from that parameter count.[6]

Microsoft’s llmlingua-2-xlm-roberta-large-meetingbank (Microsoft) has 559M parameters and 2.2 GB of F32 weights.[4] Its 4-bit weight memory is 279.5 MB, estimated (params x 0.5 bytes) from that parameter count.[4]

OpenAI’s privacy-filter (OpenAI) has 1.4B parameters, with published BF16 weights listed at 5.6 GB.[1] Its 4-bit weight memory is 700 MB, estimated (params x 0.5 bytes) from the parameter count.[1] A comparable weight-memory figure is unavailable for NVIDIA’s gliner-PII (NVIDIA).[5]

Treat the calculated values as weight-storage estimates, not total application memory requirements. Parameter arithmetic alone does not establish quantized checkpoint availability, inference memory use, or whether a model will fit on a particular device.

Which local named entity recognition models support long contexts?

LiquidAI’s LFM2.5-Encoder-350M-PII-Detector and OpenAI’s privacy-filter support long contexts, with a stated context length of 128K tokens each.[6][1] Both are candidates when your entity-detection workflow needs a long input window. Matching context limits do not establish matching detection accuracy: both have no public benchmark yet.

LiquidAI’s LFM2.5-Encoder-350M-PII-Detector was first published on Hugging Face on June 12, 2026, and uses the lfm1.0 license.[6] Its parameter count is 355M; 4-bit weight memory is approximately 177.5 MB, estimated (params x 0.5 bytes).[6] OpenAI’s privacy-filter was first published on April 17, 2026, and uses Apache-2.0.[1] Its parameter count is 1.4B; 4-bit weight memory is approximately 0.7 GB, estimated (params x 0.5 bytes).[1] Those calculations cover weights alone and do not establish total runtime memory or hardware fit.

Microsoft’s llmlingua-2-bert-base-multilingual-cased-meetingbank has a context length of 512 tokens.[2] Microsoft’s llmlingua-2-xlm-roberta-large-meetingbank supports 514 tokens.[4] Both belong to the prompt-compression family described in Microsoft’s paper, “LLMLingua-2: Data Distillation for Efficient and Faithful Task-Agnostic Prompt Compression.”[3]

For a long-context shortlist, start with the LiquidAI and OpenAI models. Choose between them after checking license requirements and evaluating detection quality on representative documents; the shared context limit alone does not identify an accuracy winner.

What licenses apply to local named entity recognition models?

Local named entity recognition models in this selection use Apache-2.0, MIT, nvidia-open-model-license, or lfm1.0, depending on the exact checkpoint.[1][2][4][5][6] Treat licensing as a checkpoint-level choice when selecting a model for a local deployment.

OpenAI’s privacy-filter (OpenAI) and Microsoft’s llmlingua-2-bert-base-multilingual-cased-meetingbank (Microsoft) list Apache-2.0.[1][2] Microsoft’s llmlingua-2-xlm-roberta-large-meetingbank (Microsoft) lists MIT.[4] The Microsoft checkpoints therefore have different license labels despite belonging to the same LLMLingua-2 family.[2][4] Record the full repository name alongside the license in your dependency inventory so that a checkpoint substitution also triggers a license review.

NVIDIA’s gliner-PII (NVIDIA) lists nvidia-open-model-license.[5] Liquid AI’s LFM2.5-Encoder-350M-PII-Detector (LiquidAI) lists lfm1.0.[6] Keep those exact identifiers in your deployment documentation, and review the corresponding terms against your intended use before adopting either checkpoint.

For an engineering handoff, include the selected repository, its license identifier, and the accompanying license text. Review your planned use, modification, and redistribution against that text. A local deployment describes where inference runs; keep the licensing decision explicit in the same way you document the model dependency. Recheck the license when changing checkpoints, including substitutions within a model family.

Frequently Asked Questions

How are the local named entity recognition models ranked?

Ranking note: only token classification models from labs with a published paper or leaderboard record qualify, using the Hugging Face token-classification pipeline tag. No public benchmark scores any candidate, so the order is newest release first, then downloads: LiquidAI’s LFM2.5-Encoder-350M-PII-Detector (LiquidAI) [6], OpenAI’s privacy-filter (OpenAI) [1], NVIDIA’s gliner-PII (NVIDIA) [5], Microsoft’s llmlingua-2-bert-base-multilingual-cased-meetingbank (Microsoft) [2], and Microsoft’s llmlingua-2-xlm-roberta-large-meetingbank (Microsoft) [4].

Why does LiquidAI lead, and why are older models included?

LFM2.5-Encoder-350M-PII-Detector (LiquidAI) ranks first because its June 12, 2026 publication is the newest among the listed candidates. [1][2][4][5][6] The family has too few releases within the last twelve months, so the newest available candidates are included. NVIDIA’s gliner-PII was published on August 29, 2025, while both Microsoft LLMLingua-2 candidates were published on March 17, 2024. [2][4][5] The ranking therefore reflects release recency, not demonstrated accuracy.

How much memory would quantized weights require?

LiquidAI’s detector has 355M parameters [6]: 0.1775 GB estimated (params x 0.5 bytes). OpenAI’s privacy-filter has 1.4B parameters [1]: 0.7 GB estimated (params x 0.5 bytes). Microsoft’s BERT candidate has 177M parameters [2]: 0.0885 GB estimated (params x 0.5 bytes). Microsoft’s XLM-RoBERTa candidate has 559M parameters [4]: 0.2795 GB estimated (params x 0.5 bytes). Each figure describes theoretical weight storage, not a measured runtime footprint or verified quantization support.

Which candidates offer long context windows?

LFM2.5-Encoder-350M-PII-Detector (LiquidAI) and OpenAI’s privacy-filter each list a context length of 128K tokens. [6][1] Microsoft’s BERT candidate lists 512 tokens, while its XLM-RoBERTa candidate lists 514 tokens. [2][4] For document processing, use those limits when planning input segmentation. A context specification alone does not establish entity extraction accuracy, and no public benchmark yet compares these candidates.

What licenses do the local candidates use?

OpenAI’s privacy-filter and Microsoft’s BERT candidate use Apache-2.0. [1][2] Microsoft’s XLM-RoBERTa candidate uses MIT. [4] NVIDIA’s gliner-PII uses nvidia-open-model-license, while LFM2.5-Encoder-350M-PII-Detector (LiquidAI) uses lfm1.0. [5][6] Check the applicable license text against your intended deployment before choosing a model. Treat license suitability as a separate decision from release date, context length, or parameter count.

Are there benchmark results proving which candidate extracts entities accurately?

Every listed candidate has the same benchmark status: no public benchmark yet. No dated benchmark result establishes an accuracy winner across this shortlist. Microsoft’s candidates are associated with “LLMLingua-2: Data Distillation for Efficient and Faithful Task-Agnostic Prompt Compression,” published on March 19, 2024. [3] Prompt compression is the paper’s stated task; its publication date is not an entity recognition benchmark date. Evaluate extraction quality on your own labeled examples.

Sources

1. openai/privacy-filter model card (Hugging Face) — 2026-09-25
2. microsoft/llmlingua-2-bert-base-multilingual-cased-meetingbank model card (Hugging Face) — 2026-09-25
3. LLMLingua-2: Data Distillation for Efficient and Faithful Task-Agnostic Prompt Compression — 2024-03-19
4. microsoft/llmlingua-2-xlm-roberta-large-meetingbank model card (Hugging Face) — 2026-09-25
5. nvidia/gliner-PII model card (Hugging Face) — 2026-09-25
6. LiquidAI/LFM2.5-Encoder-350M-PII-Detector model card (Hugging Face) — 2026-09-25