Latest30 Sep 2026Embedding/other
Embed 5 (embed-v5.0-pro, embed-v5.0-fast)
Cohere's fifth-generation embedding family for search and retrieval, in Pro and Fast tiers that share one embedding space. It embeds text, images and mixed text-and-image inputs in more than 100 languages, with a 128K-token context. Cohere reports 85.8 on ViDoRe V3 for Pro.
- Text, image and fused text+image inputs
- 128K-token context; 100+ languages
- Output dimensions 256 to 2048 (Matryoshka)
- int8 and binary output formats
AccessCohere API; Model Vault; Microsoft Foundry; Amazon SageMaker; self-hosting via vLLM. Open weights not mentioned.
PricePro: $0.12 per 1M text tokens, $0.40 per 1M image tokens; Fast: $0.08 per 1M text tokens, $0.40 per 1M image tokens (input only)
10 Sep 2026Open-weights LLM
North Small Translate
A 218B-parameter mixture-of-experts machine-translation model with 25B active parameters, covering more than 50 languages. Cohere built it with RWS and released it as non-commercial open weights. In Cohere's own evaluation it scores 83.6 on WMT26, ahead of DeepL NextGen and Google Translate.
- 218B total / 25B active MoE
- 50+ languages
- 16k input and 16k output tokens
- Lab-stated WMT26 83.6 vs DeepL NextGen 81.37
AccessOpen weights on Hugging Face under CC BY-NC 4.0 (research and non-commercial use); Cohere API (north-small-translate-1-0); commercial deployment through RWS Language Weaver.
27 Aug 2026Embedding/other
Parse (parse-v5.0)
A vision-language model for high-volume document parsing. It converts tables, forms, diagrams and images into Markdown with bounding boxes, in nine major languages. Cohere reports a ParseBench average of 79.2, against 74.5 for Mistral OCR 4 in its own comparison, and throughput of 4.5 pages per second on one GPU.
- Outputs Markdown with bounding boxes
- Handles tables, forms, diagrams and images
- Nine major languages
- Lab-stated ParseBench 79.2
AccessCohere API; Model Vault; Microsoft Foundry; AWS SageMaker; private on-premises deployment. Proprietary, not open weights.
9 Jun 2026Coding model
North Mini Code
Cohere's first model for developers: a 30B-parameter mixture-of-experts coding model with 3B active parameters. It is trained for agentic software engineering and terminal tasks, runs on a single H100, and is released under Apache 2.0. Cohere reports 33.4 on the Artificial Analysis Coding Index.
- 30B total / 3B active MoE
- 256K context; 64K max generation
- Runs on one H100 GPU
- Lab-stated Artificial Analysis Coding Index 33.4
AccessOpen weights on Hugging Face (Apache 2.0); Cohere API (north-mini-code-1-0); Model Vault; OpenRouter (free).
20 May 2026Open-weights LLM
Command A+
Cohere's first mixture-of-experts model and, per Cohere, the last in the Command A family. It combines image input, reasoning, translation and agentic tool use. It has 218B total and 25B active parameters, supports 48 languages and is released under Apache 2.0. Cohere calls it its fastest and most powerful language model yet.
- 218B total / 25B active MoE
- 128K input context; 64K max output
- Text and image input; 48 languages
- Runs on 1x B200 or 2x H100 (W4A4)
- Lab-stated Artificial Analysis Intelligence Index 37
AccessOpen weights on Hugging Face (Apache 2.0; BF16, FP8 and W4A4 builds); Cohere API (command-a-plus-05-2026); Model Vault.
26 Mar 2026Audio/voice model
Cohere Transcribe
Cohere's first speech-recognition model, a 2B-parameter conformer encoder-decoder for 14 languages released under Apache 2.0. Cohere reports that it ranked first on Hugging Face's Open ASR Leaderboard at launch, with a 5.42% average word error rate, ahead of Whisper Large v3.
- 2B parameters, conformer encoder-decoder
- 14 languages incl. English, Chinese, Japanese, Arabic
- Lab-stated 5.42% average WER (Open ASR Leaderboard)
- Apache 2.0 open weights
AccessOpen weights on Hugging Face (Apache 2.0); free rate-limited Cohere API; Model Vault (paid dedicated instances).
17 Feb 2026Small/fast LLM
Tiny Aya
A family of 3.35B-parameter open-weight multilingual models from Cohere Labs that covers more than 70 languages and runs on laptops without an internet connection. It comes as a global model plus three regional variants (Earth, Fire and Water), each specialised for a different group of languages.
- 3.35B parameters
- 70+ languages
- 8K-token context
- Global plus three region-specialised variants
AccessOpen weights on Hugging Face (CohereLabs, CC-BY-NC non-commercial); Cohere Platform; Kaggle and Ollama (per TechCrunch). Reasoning variants Tiny Aya L2-Thinker and Tiny Aya En-Thinker, with a 32K-context Tiny Aya Base 32K, released on Hugging Face 2026-10-06 (gated, CC-BY-NC).
11 Dec 2025Embedding/other
Rerank 4 (rerank-v4.0-pro, rerank-v4.0-fast)
Cohere's fourth-generation reranking models for enterprise search and retrieval. Pro is tuned for quality and Fast for low latency and high throughput. Both are multilingual and take a 32K-token context, up from 4K for Rerank 3.5. Cohere calls Rerank 4 its most powerful reranker yet.
- Pro and Fast tiers
- 32K-token context
- Multilingual reranking
AccessCohere API; Model Vault.