Run locally

📊 Small-scale; hybrid think/non-think; punches above size on reasoning

📊 MMLU 49.3, IFEval 59.5, GSM8K 44.4

📊 Text-only; solid small-model chat

📊 Capable tiny model for size

📊 Strong math reasoning for 1.5B (AIME/MATH); CoT traces

📊 Strong tiny on-device chat; good IFEval for size

📊 Strong for size on math/reasoning vs Qwen2.5-3B

📊 Compact enterprise model; thinking mode + FIM

📊 Efficient bilingual small model for edge

📊 AIME 2025 47.9; outperforms comparable-size reasoners on math

📊 Strong at 3B-4B scale; dual-mode reasoning, 6 languages, long context

📊 MMLU 63.4, IFEval 77.4, GSM8K 77.7, HumanEval ~50

📊 Strong small model via pruning+distillation

📊 Improved instruction following + tool calling; 12 languages

📊 MMLU ~69, strong reasoning for 3.8B, 128K context

📊 Rivals Qwen2.5-72B-Instruct on several tasks (Qwen claim)

📊 Strong multilingual + reasoning for 3.8B; function calling

📊 Math-focused; distilled from DeepSeek-R1 synthetic math data

📊 Solid small bilingual chat

📊 SOTA-class under 13B at release; strong math/reasoning

📊 MMLU ~62, classic strong 7B baseline, function calling

📊 AIME/MATH strong for 7B; outperforms many non-reasoning models

📊 Beats Mistral 7B and Llama 3.1 8B on many tasks; 128K context

📊 Strong bilingual EN/KO instruction-following

📊 AIME 2025 59.6; strong math/science/coding reasoning for size

📊 MMLU 69.4, HumanEval 72.6, GSM8K 84.5, IFEval 80.4

📊 Strong CoT math/reasoning for 8B

📊 Enterprise-tuned; thinking mode, FIM, strong RAG/tool use

📊 Surpasses Llama3.1-8B and Qwen2.5-7B on reasoning/knowledge tasks

📊 Top-tier speed/quality for 7B; excels at RAG, tool use, agents; 23 languages

📊 MMLU ~77, strong math/code; hybrid thinking

📊 Strong bilingual (EN/ZH) chat; competitive ~9B coding/math

📊 Beats Llama-3-8B on semantics/math/reasoning/code/knowledge; 26 languages

📊 MMLU ~71, beat Llama-3-8B on many tasks at release

📊 Best-in-class under 13B at release (depth up-scaled from 7B)

📊 Approaches o1-mini on reasoning; strong AIME/MATH

📊 MMLU 84.8, GPQA 56.1, MATH 80.4, HumanEval 82.6

📊 AIME 2024 75.3, HumanEval+ 92.9, IFEval 83.4, OmniMath 76.6

📊 AIME 2024 81.3, AIME 2025 82.5, HumanEval+ 92.3, IFEval 84.9, OmniMath 81.9

📊 GPQA ~60s, strong AIME/LiveCodeBench for size

📊 ~OpenAI o3-mini level on core reasoning; strong tool use / function calling

📊 Comparable to much larger models; improved instruction following, function calling, less repetition vs 3.1

📊 MMLU ~75, strong text chat at release

📊 Flagship-level coding in a 27B dense model (Qwen3.6 release); strong agentic coding + thinking preservation

📊 MMLU-Redux 89.3, GPQA 70.4, AIME25 70.9, LiveCodeBench v5 62.6 (2507 update)

📊 Hybrid Mamba-2 + attention; efficient long-context enterprise tasks

📊 Powerful bilingual EN/KO performance at 32B

📊 AIME 2024 90.0; matches DeepSeek-R1 (671B) on AIME 2025

📊 Flagship dense Qwen3; competitive with much larger models on reasoning/code

📊 Outperforms OpenAI o1-mini; SOTA dense reasoning at release

📊 Competitive with DeepSeek-R1 on math/reasoning despite 32B size

📊 Competitive with much larger models on chat/reasoning at release

📊 HumanEval 61.6, MBPP 52.4

📊 HumanEval 65.2, MBPP 49.4

📊 HumanEval ~41.5, LiveCodeBench ~12; 52 languages, 128K context

📊 HumanEval 70.7, MBPP 69.2

📊 HumanEval 72.5 (HumanEval+ 67.7), MBPP 72.7, BigCodeBench 33.3, LiveCodeBench 12.8

📊 HumanEval 31.1, MBPP 43.6 (base, code-completion focused)

📊 HumanEval ~31.7, strong FIM; trained on The Stack v2 (17 langs, 3T+ tokens)

📊 HumanEval 84.1, MBPP 73.6

📊 HumanEvalSynthesize ~exceeds CodeLlama-34B-Instruct; enterprise RAG/tool-use tuned

📊 HumanEval 78.6, MBPP 65.4, DS-1000 strong

📊 HumanEval 34.8 (instruct), MBPP ~44

📊 HumanEval 35.4, trained on The Stack v2 (17 langs, 3.5T+ tokens)

📊 HumanEval 75.0, beats CodeGemma-1.1-7B (61) and DeepSeek-v1.5-7B (66)

📊 HumanEvalSynthesize Python 62.2 (avg 51.4), MBPP solid; 116 languages

📊 HumanEval 88.4, MBPP 83.5, Aider ~57

📊 HumanEval 83.5 (HumanEval+ 78.7), MBPP 79.1, BigCodeBench 40.3, LiveCodeBench 23.2

📊 HumanEval 60.4 (v1.1; 56.1 v1.0), MBPP 55.2

📊 HumanEval 85.4, MBPP 73.8, LiveCodeBench 23.4 (only sub-10B model above 20%)

📊 HumanEval 36.0 (base; instruct ~42.7), MBPP ~49

📊 HumanEval 81.1, MBPP+ 68.8, supports 338 languages

📊 HumanEval 89.6, MBPP 86.2, Aider ~62

📊 HumanEval 46.3, MBPP ~66; trained on 600+ languages, 4T+ tokens

📊 HumanEval 72.6 (surpasses CodeLlama-70B-Instruct's 72.0); fully self-aligned, no GPT distillation

📊 HumanEvalSynthesize avg ~mid-30s; outperforms 2x-larger CodeLlama on instruct tasks

📊 HumanEval 81.1, MBPP 78.2, 80+ languages, native fill-in-the-middle

📊 HumanEvalSynthesize avg 41.9 (best of Granite-Code, near CodeLlama-70B-Instruct's 41.1)

📊 HumanEval 79.3, MBPP 70.0; beats CodeLlama-34B by ~8pts, ~GPT-3.5-turbo level

📊 HumanEval 53.7 (base; instruct ~50), on par with original ChatGPT/GPT-3.5

📊 HumanEval 92.7, MBPP 90.2, Aider 73.7, LiveCodeBench 31.4

📊 MMMU 29.0, DocVQA 58.3, OCRBench 52.6, TextVQA 49.9, Video-MME 33.7; smallest VLM in the world

📊 MMMU 33.7, DocVQA 70.5, Video-MME 42.2; near-2B quality at a fraction of size

📊 MMMU ~43, DocVQA ~88, strong OCR; InternVL3 family tops out at MMMU 72.2 (78B)

📊 MMMU ~48, DocVQA ~89, ChartQA strong; HallusionBench improved over 2.5

📊 VQAv2 78.1, GQA 59.0, TextVQA 44.1, DocVQA (newer builds) ~70; punches at 7B level for size

📊 MMMU ~57, strong OCR (32 languages), DocVQA ~92; current-gen (2025) successor to Qwen2.5-VL

📊 MMMU 43.0, MMBench 81.9, TextVQA 72.0, multi-frame/video summarization; 128K context

📊 MMMU ~52, DocVQA ~91, OCRBench strong

📊 MMMU 53.1, DocVQA ~93, OCRBench strong

📊 MMMU ~39, DocVQA ~73, TextVQA strong; 128K context, 140+ languages

📊 MMMU ~63, DocVQA ~94, ChartQA strong, 32-language OCR; current-gen 2025

📊 DocVQA, ChartQA, AI2D, OCRBench rival/beat Llama 3.2 11B Vision & Pixtral 12B on enterprise doc tasks; tuned for visual document understanding

📊 MMMU 35.3, improved OCR/chart reading vs LLaVA-1.5; dynamic hi-res tiling up to 672x672

📊 MMMU 42.0, DocVQA 80.0, TextVQA strong, Video-MME 52.1; best memory efficiency in class

📊 OpenCompass ~65, MMMU ~49, DocVQA ~90, OCRBench ~85 (SOTA among small models); GPT-4V-level multi-image & video

📊 MMMU ~73, DocVQA 92.7, ChartQA/InfoVQA strong, OCRBench ~88; among best open 8B VLMs

📊 MMMU 58.6, DocVQA 95.7, ChartQA ~87, OCRBench ~86; beats Llama 3.2 11B Vision on most VQA

📊 MMMU ~69, DocVQA ~95, MathVista strong; Qwen3-VL family scores up to MMMU 80.6 at largest sizes

📊 MMMU 50.7, DocVQA 88.4, ChartQA ~83, AI2D ~91, VQAv2 ~75

📊 MMMU ~36, MMBench ~70, better text-in-image reading than 1.5

📊 MMMU 52.5 (CoT), DocVQA 90.7 (ANLS), ChartQA ~82, VQAv2 ~78; strong multi-image

📊 MMMU 50.3, DocVQA 82.3, InfoVQA/ChartQA strong; 128K context

📊 MMMU 56.1, DocVQA 85.6, ChartQA/AI2D strong; competitive with much larger models

📊 MMMU ~73-75, DocVQA ~96; MoE with only ~3B active params so runs fast

📊 MMMU ~46, MMBench ~79; strongest LLaVA-NeXT tier

📊 MMMU ~70, DocVQA ~94, MathVista strong; near 72B quality

📊 27M params; better average WER than Whisper tiny.en while ~5x faster. v2 Tiny hits ~50ms latency (5.8x faster than Whisper Tiny). English. Variable-length input (no fixed 30s padding) = big edge speedup.

📊 61M params, 237MB on disk; beats Whisper base.en on average WER while running much faster on CPU. English. Variable-length encoder avoids Whisper's fixed-window overhead.

📊 Compact FastConformer hybrid TDT+CTC; competitive English WER for its size, very high RTFx. Good edge/streaming candidate.

📊 >1200 RTFx (extremely fast); 4 languages (en/de/fr/es) ASR + translation. Strong accuracy-per-param for a 182M model. Word-level timestamps.

📊 Non-autoregressive; >5x faster than Whisper-Small and ~15x faster than Whisper-Large; latency <80ms. Beats Whisper on Chinese/Cantonese benchmarks (e.g. AISHELL-1). 50+ languages incl. zh/en/yue/ja/ko, plus emotion (SER) + audio-event detection (AED) + ITN.

📊 Pretrained on 436k hrs across 128 languages. Not directly an ASR head — needs fine-tuning (CTC) per language; fine-tuned variants reach competitive multilingual WER (e.g. Common Voice). Foundation for many community ASR models.

📊 1.8% / 3.3% WER on LibriSpeech test-clean / test-other (CTC, self-training on 960h + 53k unlabeled). English-only, no built-in punctuation. With 10 min labeled data still ~4.8/8.2 WER.

📊 ~3-4% WER LibriSpeech test-clean; only 0.3-0.7pt WER worse than large-v2 but ~6-8x faster (4 decoder layers vs 32). 99-language multilingual.

📊 Within 1.5% WER of large-v3 on OOD short-form, within 1% on long-form, +0.1% better on chunked long-form. ~6x faster than large-v3. English-only.

📊 Short-form 7.10 WER vs large-v3's 7.14 (slightly better); long-form 10.04 vs 8.82 (a bit worse). ~1.5x faster than large-v3-turbo on long-form. Trained on 98k hrs with patient teacher + SpecAugment.

📊 Near large-v3-turbo quality (~3-4% WER LibriSpeech clean) at very high throughput; combines turbo's pruned decoder with CTranslate2 speedups. Sub-second latency feasible.

📊 Streaming STT with ~0.5s delay + semantic VAD; English & French. Word-level timestamps; robust to noise. Built on Mimi codec + Moshi-style autoregressive decoder. Trained on 2.5M hrs.

📊 Open ASR Leaderboard avg 6.05% WER (was #1 at release, May 2025). LibriSpeech test-clean 1.69%, test-other 3.19%. RTFx >3000 — transcribes ~1hr audio per second on GPU. English-only.

📊 Open ASR Leaderboard avg 6.34% WER. LibriSpeech test-clean 1.93%. Multilingual Fleurs avg 11.97% WER across 25 European languages. RTFx >3000. Auto language detection.

📊 Same accuracy as Whisper large-v3 (~1.8-2.7% WER LibriSpeech clean) but up to 4x faster and lower memory via CTranslate2. int8 quant adds speed with minimal WER loss.

📊 Avg WER ~6.67% on Open ASR Leaderboard; >1000 RTFx (much faster than original Canary-1B). 4 languages (en/de/fr/es) ASR + En<->X translation with optional punctuation/capitalization.

📊 Top-tier on Open ASR Leaderboard (~5.6-6.7% WER region); 25 European languages, ASR + speech translation (X<->En). Encoder-decoder FastConformer + Transformer. Word-level timestamps.

📊 ~6.0-6.5% avg WER region on Open ASR Leaderboard; trained on 64k+ hrs. Larger encoder than 0.6B for marginal accuracy gains. English.

📊 ~6.4% WER; English-only, optimized for max accuracy with a 2.5s delay. Robust in noisy conditions and on audio up to ~2 hours. A H100 can serve ~400 streams in real-time.

📊 No published MOS; positioned as 'SoTA under 25MB'. 24kHz output, 8 voices. Real-time on CPU including phones/Raspberry Pi.

📊 No formal MOS; VITS-based; ~10x real-time on desktop CPU, real-time on Raspberry Pi 5. Medium voices 22.05kHz, high 22.05kHz.

📊 VITS-based; CPU real-time capable. No formal MOS published but widely used; clear, natural multilingual speech. ~44.1kHz internal.

📊 Was #1 in TTS Spaces Arena (Dec 2024) at only 82M params, beating much larger models on naturalness ELO. ~24kHz. Sub-real-time on CPU, very fast on GPU.

📊 Tone-color conversion step <100ms; instant zero-shot voice cloning. Quality inherits from MeloTTS base. No single MOS, but strong cross-lingual cloning fidelity.

📊 LJSpeech MOS-N 4.55 vs 4.23 ground-truth (surpasses human recordings single-speaker); matches human on multispeaker VCTK. 24kHz. Diffusion-style prosody.

📊 Flow-matching (non-autoregressive, no diffusion) -> fast inference + strong prosody. ~0.15-0.3 RTF on GPU. Excellent zero-shot cloning + code-switching; among the top open cloning models of 2024-25. 24kHz.

📊 Vendor blind test: 65.3% preferred Chatterbox-Turbo vs 24.5% ElevenLabs (take with salt). Sub-200ms latency (Turbo ~472ms first chunk, RTF ~0.5). First open model with emotion-exaggeration control. Trained on 500K hrs.

📊 Conversational/contextual prosody (uses prior turns of text+audio). Llama backbone + Mimi RVQ decoder. ~200ms-class streaming. Strong context-aware naturalness; no single MOS published.

📊 6-second zero-shot voice cloning, 17 languages, cross-lingual + emotion/style transfer, 24kHz. Long the community favorite for quality cloning; ~150-200ms streaming latency on GPU.

📊 Llama-3.2-3B Speech-LLM, trained 100K+ hrs English. ~200ms streaming latency (down to ~100ms with input streaming). Zero-shot cloning + inline emotion tags (<laugh>,<sigh>,<gasp>...). Claims to rival/surpass closed-source naturalness.

📊 Trained on 45K hours. Natural, controllable speech; you steer gender/pitch/pace/reverb/emotion with a natural-language description prompt. No headline MOS but fully reproducible (data+code+weights open).

📊 Specialized for ultra-realistic multi-speaker DIALOGUE in one pass; handles nonverbals (laughs, coughs, throat-clear). Real-time on enterprise GPUs (~40 tok/s on A4000). 44.1kHz. Audio-conditioned emotion/tone + voice cloning from <=10s clip.

📊 Smallest Arctic; dim 384

📊 MTEB (English v1) ~56.3 avg; the classic fast/CPU baseline

📊 Fast English retrieval, tiny footprint

📊 MTEB (English v1) ~61.4 avg

📊 MTEB (English v1) ~59.9 avg

📊 Compact English retrieval, dim 384

📊 MTEB (English v1) ~62.2 avg — punches above its size

📊 MTEB retrieval ~54.9 nDCG@10 (English)

📊 MTEB (English v1) ~63.5 avg

📊 MTEB (English v1) ~61.5 avg

📊 MTEB (English v1) ~57.8 avg — long the best general-purpose ST model

📊 Good multilingual quality for 118M params

📊 Beats OpenAI text-embedding-ada-002 & 3-small on short+long context; MTEB ~62

📊 MTEB (English v1) ~64 avg

📊 2025 R2 release; improved retrieval over r1, longer context

📊 Solid multilingual MTEB, mid-size

📊 Fast multilingual cross-encoder; strong BEIR/MKQA; agentic function-calling rerank

📊 Competitive multilingual retrieval; enterprise/clean-data trained

📊 Strong multilingual retrieval, smaller footprint than L-v2.0

📊 Highest-ranked open multilingual embedder under 500M on MMTEB at release (Sep 2025)

📊 MTEB Multilingual mean 64.33; MTEB-Code strong; instruction-aware

📊 Cross-encoder reranker; strong MTEB-R / MIRACL reranking gains; instruction-aware

📊 MTEB retrieval ~55.9 nDCG@10 (English) at release (Apr 2024)

📊 MTEB (English v1) ~64.2 avg; long the default RAG baseline

📊 MTEB (English v1) ~62.3 avg

📊 MTEB (English v1) ~64.7 avg — SOTA for BERT-large size at release (Mar 2024), no MTEB-data overlap

📊 MTEB (English v1) ~65 avg — SOTA in its size class at release

📊 MTEB (English v1) ~70 avg — top small model; near 1.5B quality

📊 Multilingual MoE; competitive multilingual MTEB at ~305M active params

📊 XLM-RoBERTa-base cross-encoder; solid CN/EN reranking

📊 Strong multilingual MTEB; beats BGE-large-en & Cohere multilingual-v3 at release

📊 Outperforms OpenAI text-embedding-3-large & Cohere on MTEB multilingual at release (Sep 2024)

📊 MIRACL nDCG@10 ~70 (multilingual SOTA at release); strong BEIR; hybrid dense+sparse+ColBERT

📊 Multilingual cross-encoder; strong MIRACL/BEIR reranking; lightweight

📊 Top BEIR nDCG@10 + strong CLEF/MIRACL multilingual at release (Dec 2024)

📊 MTEB Multilingual mean 69.45; near-SOTA retrieval

📊 SOTA-class open reranker; large gains on BEIR/MIRACL/MTEB-R reranking

📊 MTEB ~67 avg; instruction-tuned LLM embedder

📊 SOTA-class open reranker (2025); strong BEIR; Qwen2.5-1.5B backbone

📊 MTEB (English v1) ~71.2 avg — top-tier open English embedder; basis of jasper (MTEB #2)

📊 MTEB Multilingual mean 70.58 — #1 on MTEB multilingual leaderboard (Jun 5 2025)

📊 Best open reranker quality in Qwen3 series; top BEIR/MIRACL reranking

📊 MTEB ~70 avg — #1 English & Chinese MTEB at release (Jun 2024)