1
Cohere Command is a family of generative language models for building enterprise agents, assistants, RAG applications, and text-generation workflows. The models are available through APIs and private deployment options.
Best Language Models 2026
Language models are open-weight or proprietary generative models offered as downloadable weights or through an inference application programming interface (API). Engineers and architects build them into applications and agents rather than buying a vendor's finished chat application.
We’ve collected videos, features, and capabilities below. Take me there.
All Products
Learn More about Language Models Software
What Are Language Models?
Language models are open-weight or proprietary generative models offered as downloadable weights or through an inference application programming interface (API). Engineers and architects build them into applications and agents rather than buying a vendor's finished chat application.
The core test is whether the buyer is selecting a model to put into software or buying an application that contains a model. This category covers large language models (LLMs) and small language models (SLMs) that generate text or code from a prompt, including multimodal models accepting image, audio, or video inputs. Teams use them in retrieval-augmented generation (RAG) pipelines, agents, and backend services while keeping control over integration and workflow design.
Teams choose smaller, open, or embeddable models over the largest frontier model when a bounded task such as extraction, classification, or RAG synthesis meets its quality target with a smaller model. Downloadable weights let a team choose where the model runs (without by itself establishing data-residency compliance); compact models can improve response time and hardware fit on a suitable runtime (including CPU and on-device); open weights let a team inspect and select license terms, pin versions, and, where the license permits, fine-tune and deploy on its own data; and predictable cost at volume depends on serving design and utilization. The trade-off is operational work: the team takes on evaluation and deployment, and a smaller or self-hosted model is not automatically cheaper or better.
How Language Models Differ from Related Software
- Conversational AI covers finished chat apps for end users, not models built into software.
- AI Model Serving & Inference covers runtime engines and hosts (such as vLLM, Ollama, Together AI, and Amazon Bedrock) that execute weights, rather than the models.
- AI Gateways route and govern traffic across model providers.
- AI Infrastructure & Development and LLMOps supply training, tuning, evaluation, and observability tools.
- Autonomous AI Agents are task-executing systems that call models and tools.
- Standalone retrieval and single-modality models, such as embedding, reranking, speech, or image models, fall outside this generative scope.
Language Model Features
- Text, code, and multimodal generation - Generates text or code from prompts, with select models accepting image, audio, or video inputs and long contexts.
- Weights or API access - Supplies the model as downloadable weights under published license terms, through an inference API, or both.
- Architecture and latency options - Includes compact dense models (such as Microsoft Phi-4-mini-instruct at 3.8B parameters), quantized weights for CPU or on-device use, and sparse Mixture-of-Experts (MoE) designs activating a subset of parameters per token to cut latency.
- Structured outputs and tool calling - Produces schema-adherent JSON and function calls for APIs, databases, and Model Context Protocol (MCP) workflows.
- Configurable reasoning controls - Exposes explicit reasoning modes or budgets (as in Nous Research Hermes 4 14B, IBM Granite 4.2, and Upstage Solar Pro 4); hiding reasoning output is separate from switching computation off.
- Fine-tuning and portability - Where weights and license terms permit, teams can tune models on domain data with Low-Rank Adaptation (LoRA) and move them across compatible engines.
How to Choose a Language Model
- Task evaluation - Public leaderboards rarely predict production RAG grounding or schema adherence. Test candidates on representative prompts, retrieval passages, and latency targets.
- Architecture and memory - A low active-parameter count in a sparse MoE model does not imply a small total memory requirement: weights, precision, runtime, and key-value (KV) cache all count, though weights can be offloaded across GPU, CPU, and disk with performance trade-offs.
- License, hosting, and filtering - Downloadable weights are not automatically open-source under the Open Source Initiative (OSI) definition. Apache 2.0 models (Google Gemma 4, IBM Granite 4.2, Mistral Small 4) and MIT models (Microsoft Phi-4-mini-instruct) carry conditions such as attribution, while custom licenses (such as Mistral Large 4's Mistral Ecosystem Open Weight License) add commercial limits. License obligations, hosting location, and provider filtering are separate questions; self-hosting shifts compliance to the deploying team.
- Lifecycle and version pinning - Providers deprecate API snapshots and update checkpoints and tokenizers. Pin versions and run regression tests before upgrades.
Pricing Information
Costs depend on metered API token volume (input, output, cached prompts, and batch rates), prepaid tiers, or compute for weights. Standard list rates per 1M tokens (as of November 2026) include:
- Cohere Command - $0.0375 input and $0.15 output for Command R7B, and $2.50 input and $10.00 output for Command A (distinct from open-weight Command A+).
- Upstage Solar - $0.10 input and $0.40 output for Solar Mini 4, and $0.30 input and $1.20 output for Solar Pro 4 (which has an active promotion with no end date).
- Mistral AI - $0.15 input and $0.60 output for Mistral Small 4, and $0.70 input and $2.10 output for Mistral Large 4.
- Amazon Nova on Amazon Bedrock - For generative language models in the Nova family, Nova 2 Lite is $0.30 input and $2.50 output.
Many open-weight releases can be downloaded without a license fee, but running them requires suitable compute and memory, and vendor-managed private hosting and support may require a custom quote.
Loading related categories...
Language Models FAQs
What does a language model do, and how does it differ from a chatbot application?
A language model is a generative software component that takes a prompt and produces text or code, with select models also accepting image, audio, or video inputs. A chatbot application is a finished product with a user interface, conversation state, and vendor-managed workflow rules sold to end users. A language model is the underlying engine that developers call through an application programming interface (API) or run from downloadable weights inside custom software, retrieval-augmented generation (RAG) pipelines, and agents.
What is the difference between a large and a small language model?
Large language models (LLMs) and small language models (SLMs) share generative architectures but target different deployment and workload trade-offs, with no fixed parameter cutoff across the industry. Compact dense models such as Microsoft Phi-4-mini-instruct (3.8B parameters) or Mistral's Ministral 3 3B, along with sparse Mixture-of-Experts (MoE) models such as Arcee Trinity Mini (26B total and 3B active parameters), are built to lower latency and fit single GPUs, CPUs, or on-device memory for bounded tasks like classification, extraction, and RAG synthesis. Larger models such as Mistral Large 4 or Cohere Command A tend to hold broader multi-domain knowledge and handle more open-ended synthesis, at the cost of higher memory, compute, and per-token spend.
Are open-weight and open-source language models the same?
No. An open-weight language model publishes downloadable parameter checkpoints so teams can inspect, quantize, fine-tune, and self-host the model, often under a custom license with commercial or attribution conditions. Under the Open Source Initiative (OSI) Open Source AI Definition, an open-source language model (such as Ai2 Olmo 3) does not require a permissive-only license (share-alike terms are permitted) and does not promise reproduction of an identical model; instead, it requires the preferred form for modification under OSI-approved terms, consisting of the model parameters, the complete code used to train and run the system, and sufficiently detailed data information for a skilled person to build a substantially equivalent system.
Can a language model be used through an API, or must it run locally?
Language models can be called through a hosted inference API, deployed from downloadable weights on local or private cloud hardware, or offered through both routes. API-served models (such as Upstage Solar Pro 4 or the generative language models in the Amazon Nova family) let teams ship features without managing model servers, while downloadable weights (such as Google Gemma 4, IBM Granite 4.2, or Microsoft Phi-4-mini-instruct) let organizations run inference on their own cloud instances, on-premises servers, CPUs, or edge devices.
How can a team evaluate a language model on its own data?
To apply the selection criteria in the category description above, teams should build an evaluation set from representative production inputs rather than relying on public leaderboards:
- Assemble a task-specific test set - Collect a starting pilot set of 50 to 200 real prompts, retrieval passages, and edge cases covering the target extraction, classification, tool-calling, or synthesis task.
- Score quality and failure modes - Check schema adherence, citation faithfulness to supplied RAG context, refusal behavior, and formatting errors against explicit grading rubrics.
- Measure latency and memory under load - Record time-to-first-token, throughput, and memory growth at the target concurrency and context length on the intended serving path.
- Pin and regression-test versions - Lock the specific model snapshot or weight checkpoint, quantization level, and chat template so future model swaps can be tested against the same baseline.
How much does using a language model cost?
Language model costs depend on the access path, model size, and request volume rather than a per-seat subscription. Metered inference APIs charge per 1 million input and output tokens, often with lower rates for smaller models, cached prompt prefixes, and batch jobs, as well as optional prepaid commitment tiers. Many open-weight releases can be downloaded without a license fee, but self-hosting requires suitable compute and memory, storage, and engineering maintenance, while vendor-managed private hosting and support may require a custom quote.
