The AI Models Guide — What to Use for Every Task
A living guide to the AI model landscape: 50 tracked models grouped by task, weekly movement and the skills employers hire for — refreshed automatically.
11 min read
A living guide to the AI model landscape: 50 tracked models grouped by task, weekly movement and the skills employers hire for — refreshed automatically.
Quick Answer
Task-by-task guide to AI models, refreshed weekly from live data: 50 tracked models across LLMs, embeddings, vision and speech plus demand from 3,751 AI job lis
Search Snapshot
50 models tracked · snapshot 2026-08-02 · 3,751 active AI job listings analysed
The right model depends entirely on what you need it to do — a text classifier, a speech transcriber and a semantic search engine are built on fundamentally different architectures. Picking by name or hype wastes compute and often produces worse results than a smaller, purpose-built alternative. This guide tracks 50 models across 8 task families, refreshed every week from Hugging Face download data, so you can match your use case to what developers are actually running in production. Download volume and job market demand are tracked side by side, giving you a practical signal on which skills and model types are worth investing in.
Work through these in order — most wrong choices come from starting at step 4.
| Model family | Use it for | Current leader |
|---|---|---|
| Text generation and LLMs | Chat assistants, coding copilots, agents, summarization, drafting and any task where the output is free-form text. | Qwen3-0.6B |
| Vision-language and multimodal | Describing images, reading documents and screenshots, answering questions about visual content and combining text with other media. | gemma-4-26B-A4B-it |
| Embeddings, search and ranking | Semantic search, RAG retrieval, deduplication, clustering and recommendation. Rerankers refine an initial result list for precision. | all-MiniLM-L6-v2 |
| Language understanding and classification | Sentiment analysis, topic and intent classification, named-entity recognition, translation and other structured judgements about text. | bert-base-uncased |
| Speech and audio | Transcription, voice interfaces, dubbing, meeting notes and audio event detection. | Kokoro-82M |
| Computer vision | Classifying images, detecting and segmenting objects, visual quality control and content moderation. | clip-vit-base-patch32 |
| Time series forecasting | Demand forecasting, anomaly detection and other predictions over numeric sequences and tables. | chronos-2 |
| Other specialist models | Niche tasks — protein folding, robotics, graph learning and research architectures that have not settled into a mainstream category. | electra-base-discriminator |
Use it for: Chat assistants, coding copilots, agents, summarization, drafting and any task where the output is free-form text.
Text generation and LLMs are dominated by compact, efficient models this week, with Qwen/Qwen3-0.6B leading at 28.3M downloads — well ahead of facebook/opt-125m at 17.3M and the testing-focused trl-internal-testing/tiny-Qwen2ForCausalLM-2.5 at 15.5M. The top three are all sub-billion-parameter models, signalling strong demand for lightweight deployments over large-scale alternatives.
| Model | Maker | Task | Downloads | Likes | This week |
|---|---|---|---|---|---|
| Qwen3-0.6B | Qwen | text-generation | 28.3M | 1.5k | +1.6M |
| opt-125m | text-generation | 17.3M | 286 | +1.3M | |
| tiny-Qwen2ForCausalLM-2.5 | trl-internal-testing | text-generation | 15.5M | 18 | +1.9M |
| Qwen3-8B | Qwen | text-generation | 15.1M | 1.3k | — |
| Qwen2.5-1.5B-Instruct | Qwen | text-generation | 13.9M | 786 | +538.9k |
| gpt2 | openai-community | text-generation | 13.4M | 3.4k | — |
The largest and fastest-moving family. Frontier models (Claude, GPT, Gemini) are API-only; the open-weight models listed here are what you run yourself.
Use it for: Describing images, reading documents and screenshots, answering questions about visual content and combining text with other media.
Vision-language and multimodal models are currently led by Google's Gemma 4 family, with google/gemma-4-26B-A4B-it reaching 12.1M downloads and google/gemma-4-31B-it close behind at 11.8M. Qwen/Qwen3.5-9B also sits at 11.8M downloads, showing that mid-size open models are competing directly with Google's releases in this category.
| Model | Maker | Task | Downloads | Likes | This week |
|---|---|---|---|---|---|
| gemma-4-26B-A4B-it | image-text-to-text | 12.1M | 1.3k | — | |
| gemma-4-31B-it | image-text-to-text | 11.8M | 3.4k | — | |
| Qwen3.5-9B | Qwen | image-text-to-text | 11.8M | 1.8k | +2.4M |
| Qwen2.5-VL-7B-Instruct | Qwen | image-text-to-text | 9.1M | 1.7k | — |
| Qwen3.6-35B-A3B-FP8 | Qwen | image-text-to-text | 8.3M | 334 | New |
Use these when the input is not just text — document extraction, chart reading and UI understanding all live here.
Use it for: Semantic search, RAG retrieval, deduplication, clustering and recommendation. Rerankers refine an initial result list for precision.
Embeddings, search and ranking remain among the most downloaded model types on the hub, with sentence-transformers/all-MiniLM-L6-v2 recording a commanding 251.1M downloads. Cross-encoder/ms-marco-MiniLM-L6-v2 follows at 86.8M and BAAI/bge-small-en-v1.5 at 69.7M, confirming that small, fast retrieval models dominate production search pipelines.
| Model | Maker | Task | Downloads | Likes | This week |
|---|---|---|---|---|---|
| all-MiniLM-L6-v2 | sentence-transformers | sentence-similarity | 251.1M | 5.2k | — |
| ms-marco-MiniLM-L6-v2 | cross-encoder | text-ranking | 86.8M | 293 | +1.6M |
| bge-small-en-v1.5 | BAAI | feature-extraction | 69.7M | 520 | +3.2M |
| paraphrase-multilingual-MiniLM-L12-v2 | sentence-transformers | sentence-similarity | 58.6M | 1.3k | +8.9M |
| bge-m3 | BAAI | sentence-similarity | 34.6M | 3.3k | — |
| all-mpnet-base-v2 | sentence-transformers | sentence-similarity | 25.9M | 1.3k | — |
The workhorse family behind almost every RAG system. Small, cheap to run and rarely the bottleneck — pick one with strong retrieval benchmarks and move on.
Use it for: Sentiment analysis, topic and intent classification, named-entity recognition, translation and other structured judgements about text.
Language understanding and classification is still anchored by google-bert/bert-base-uncased, which logs 104.9M downloads this period. google-t5/t5-small at 25.2M and FacebookAI/xlm-roberta-base at 21.9M hold second and third place, showing that encoder-only and seq2seq baselines remain core infrastructure for NLP practitioners.
| Model | Maker | Task | Downloads | Likes | This week |
|---|---|---|---|---|---|
| bert-base-uncased | google-bert | fill-mask | 104.9M | 2.7k | +24.4M |
| t5-small | google-t5 | translation | 25.2M | 587 | — |
| xlm-roberta-base | FacebookAI | fill-mask | 21.9M | 876 | — |
| bge-reranker-v2-m3 | BAAI | text-classification | 18.9M | 1.1k | +1.1M |
| roberta-large | FacebookAI | fill-mask | 12.5M | 317 | — |
| roberta-base | FacebookAI | fill-mask | 11.5M | 632 | +2.6M |
BERT-family encoders remain heavily used in production because they are fast, cheap and deterministic to fine-tune for a fixed label set.
Use it for: Transcription, voice interfaces, dubbing, meeting notes and audio event detection.
Speech and audio is led by hexgrad/Kokoro-82M with 11.3M downloads, putting this compact TTS model ahead of coqui/XTTS-v2 at 9.0M and pyannote/speaker-diarization-3.1 at 8.6M. Download gaps between the top three are relatively narrow, reflecting active use across synthesis and speaker-analysis tasks alike.
| Model | Maker | Task | Downloads | Likes | This week |
|---|---|---|---|---|---|
| Kokoro-82M | hexgrad | text-to-speech | 11.3M | 6.6k | +1.3M |
| XTTS-v2 | coqui | text-to-speech | 9.0M | 3.7k | — |
| speaker-diarization-3.1 | pyannote | automatic-speech-recognition | 8.6M | 2.9k | +174.6k |
| whisper-large-v3-turbo | openai | automatic-speech-recognition | 8.5M | 3.2k | New |
Speech-to-text (ASR) and text-to-speech are separate model families — most voice products chain one of each around an LLM.
Use it for: Classifying images, detecting and segmenting objects, visual quality control and content moderation.
Computer vision usage centres on CLIP-based models, with openai/clip-vit-base-patch32 at 22.9M downloads and openai/clip-vit-large-patch14 at 8.9M bracketing timm/mobilenetv3_small_100.lamb_in1k at 18.1M. The MobileNetV3 figure confirms that lightweight classification backbones still see heavy production use alongside large vision-language encoders.
| Model | Maker | Task | Downloads | Likes | This week |
|---|---|---|---|---|---|
| clip-vit-base-patch32 | openai | zero-shot-image-classification | 22.9M | 992 | — |
| mobilenetv3_small_100.lamb_in1k | timm | image-classification | 18.1M | 99 | — |
| clip-vit-large-patch14 | openai | zero-shot-image-classification | 8.9M | 2.1k | — |
Zero-shot vision models classify against labels you supply at runtime — no retraining needed when categories change.
Use it for: Demand forecasting, anomaly detection and other predictions over numeric sequences and tables.
Time series forecasting is currently dominated by Amazon's Chronos ecosystem, with amazon/chronos-2 reaching 25.7M downloads and autogluon/chronos-2 adding a further 12.0M. autogluon/chronos-bolt-small sits at 13.2M, and the combined weight of these three models signals that pretrained zero-shot forecasters have become mainstream in the category.
| Model | Maker | Task | Downloads | Likes | This week |
|---|---|---|---|---|---|
| chronos-2 | amazon | time-series-forecasting | 25.7M | 385 | +11.1M |
| chronos-bolt-small | autogluon | time-series-forecasting | 13.2M | 58 | — |
| chronos-2 | autogluon | time-series-forecasting | 12.0M | 47 | +1.7M |
Foundation models for time series are a recent arrival — they forecast series they were never trained on, the way LLMs handle unseen text.
Use it for: Niche tasks — protein folding, robotics, graph learning and research architectures that have not settled into a mainstream category.
Specialist models outside core categories show significant scale, led by google/electra-base-discriminator at 53.8M downloads. lpiccinelli/unidepth-v2-vitl14 reaches 30.2M — highlighting strong demand for monocular depth estimation — while Bingsu/adetailer at 10.3M reflects continued use in image inpainting and detection workflows.
| Model | Maker | Task | Downloads | Likes | This week |
|---|---|---|---|---|---|
| electra-base-discriminator | — | 53.8M | 147 | — | |
| unidepth-v2-vitl14 | lpiccinelli | — | 30.2M | 47 | +16.8M |
| adetailer | Bingsu | — | 10.3M | 755 | — |
The 2026-08-02 snapshot added four models to the tracked set: Qwen/Qwen3-32B, Qwen/Qwen3.6-35B-A3B-FP8, openai/whisper-large-v3-turbo and openai/gpt-oss-20b. The biggest download surge this week came from an unexpected direction — google-bert/bert-base-uncased gained 24.4M downloads, the fastest growth of any model in the snapshot, followed by depth-estimation model lpiccinelli/unidepth-v2-vitl14 at +16.8M.
| New in the tracked set | Task | Downloads |
|---|---|---|
| Qwen/Qwen3-32B | text-generation | 9.2M |
| openai/whisper-large-v3-turbo | automatic-speech-recognition | 8.5M |
| Qwen/Qwen3.6-35B-A3B-FP8 | image-text-to-text | 8.3M |
| openai/gpt-oss-20b | text-generation | 8.3M |
| Fastest download growth | Task | Added this week |
|---|---|---|
| google-bert/bert-base-uncased | fill-mask | +24.4M |
| lpiccinelli/unidepth-v2-vitl14 | — | +16.8M |
| amazon/chronos-2 | time-series-forecasting | +11.1M |
| sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2 | sentence-similarity | +8.9M |
| BAAI/bge-small-en-v1.5 | feature-extraction | +3.2M |
| FacebookAI/roberta-base | fill-mask | +2.6M |
| Qwen/Qwen3.5-9B | image-text-to-text | +2.4M |
| trl-internal-testing/tiny-Qwen2ForCausalLM-2.5 | text-generation | +1.9M |
Machine Learning appears in 1,249 active job listings, making it the single most requested skill across the 3,751 postings analysed, with LLMs and GenAI close behind at 672 listings. Embedding and language understanding models — the workhorses behind search, classification and retrieval — map directly to these demands, while the 294 listings requesting AI Agents signal growing employer appetite for orchestration skills built on top of text generation and tool-use models.
| Skill | Active listings |
|---|---|
| Machine Learning | 1249 |
| LLMs / GenAI | 672 |
| AI Agents | 294 |
| AWS | 172 |
| Azure | 142 |
| Fine-tuning | 135 |
| Deep Learning | 132 |
| Agile / Scrum | 119 |
| A/B Testing | 107 |
| GCP | 92 |
| Kubernetes | 76 |
| MLflow | 62 |
Explore the full demand data on the AI and ML job market page or track movement over time on skill trends.
Model popularity comes from a weekly snapshot of Hugging Face download and like counts (50 top models, latest snapshot 2026-08-02). "This week" columns compare the latest snapshot with one taken roughly a week earlier, so they show download velocity rather than all-time totals. Employer demand comes from daily analysis of active AI job listings (3,751 at last count). Frontier API-only models do not appear in download tables because they publish no open weights — that is a limitation of the source, not a quality judgement.
For RAG and semantic search, sentence-transformers/all-MiniLM-L6-v2 is the most widely used starting point, with 251.1M downloads making it the most downloaded model in this snapshot. If you need multilingual retrieval, sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2 added 8.9M downloads this week alone, indicating strong and active adoption. Both sit in the Embeddings, search and ranking family, which covers 13 of the 50 tracked models.
The most downloaded AI model in the current snapshot is sentence-transformers/all-MiniLM-L6-v2, with 251.1M total downloads as of 2026-08-02. It is a lightweight sentence embedding model designed for semantic similarity, clustering and search — not a large language model. Its download count reflects how heavily it is used in production retrieval pipelines rather than experimental or research contexts.
Machine Learning is the skill employers ask for most, appearing in 1,249 of the 3,751 active job listings tracked. LLMs and GenAI follow at 672 listings and AI Agents at 294, showing that generative model experience is now a distinct hiring category separate from broader ML. Fine-tuning appears in 135 listings, a practical signal that employers want engineers who can adapt existing models rather than train from scratch.
Updated August 03, 2026. This guide regenerates automatically from fresh data every week.
Continue with adjacent guides and tactical breakdowns.
Actionable guides, market updates and shipping notes — once a week.