Command A+: Unraveling Command-Centric AI

• Command A+ is an ambiguous, inferred label extending Command A's enterprise LLM framework, raising questions about its distinct existence and post-training enhancements.
• Command A's documented architecture features a 111B-parameter decoder-only Transformer with advanced techniques like hybrid attention and systematic model merging.
• Related research in drone control, user telemetry, and shell customization highlights practical command-centric applications and associated safety and deployment challenges.

Searching arXiv for papers on “Command A+” and closely related “Command A” documentation. Command A+ is not explicitly defined in the available arXiv technical literature. The nearest direct documentation concerns Command A, described as a 111B-parameter enterprise LLM from Cohere, while adjacent research uses the term “command” in distinct senses: natural-language control of MAVLink-based drones, user representation learning from command telemetry, and shell-level customization practices (Cohere et al., 1 Apr 2025, Ramos-Silva et al., 21 Jan 2026, Chu et al., 2022, Schröder et al., 2020). As a result, any precise account of Command A+ must distinguish documented facts from inference: the literature provides detailed technical evidence for Command A and for broader command-centric AI systems, but it does not state whether Command A+ exists as a distinct product, how it differs from Command A, or whether “A+” denotes a later checkpoint, stronger post-training, a deployment SKU, or a related family member (Cohere et al., 1 Apr 2025).

1. Documentary status and nomenclature

Within the supplied literature, the designation “Command A+” is an ambiguous label rather than a formally specified model name. The paper "Command A: An Enterprise-Ready LLM" states explicitly that it is about Command A and Command R7B, and that it does not mention Command A+ directly anywhere in the provided text (Cohere et al., 1 Apr 2025). This documentary gap is central to any encyclopedic treatment of the topic, because it bounds what can be asserted as fact.

The same corpus shows that “command” is a polysemous term in current research. In enterprise LLM work it refers to a model family centered on grounded generation, tool use, multilingual business tasks, and efficient deployment; in physical AI it denotes natural-language drone command and control via MCP, MAVSDK, and MAVLink; in user-modeling work it denotes telemetry streams of software commands; and in human-computer interaction it denotes shell commands and their customization through aliases (Cohere et al., 1 Apr 2025, Ramos-Silva et al., 21 Jan 2026, Chu et al., 2022, Schröder et al., 2020). This suggests that Command A+ can only be situated reliably by placing the undocumented label against these documented command-centric traditions.

A plausible implication is that Command A+ functions, in current discourse, as a referential extension of Command A rather than as a separately archived research object. That implication remains inferential, because the literature does not provide a model card, architecture table, or benchmark sheet for an entity explicitly named Command A+ (Cohere et al., 1 Apr 2025).

2. Technical substrate documented for Command A

The closest technical basis for understanding Command A+ is the published architecture of Command A. Command A is described as a decoder-only Transformer with 111B parameters, a vocabulary size of 255,000, a multilingual tokenizer with special tokens for chat turns and tool calls, SwiGLU activations, Grouped-query attention, shared input and output embeddings, no bias terms, and a parallel transformer block (Cohere et al., 1 Apr 2025). Its defining architectural feature is a hybrid attention stack with interleaved sliding-window attention and full attention layers in a 3:1 ratio; the sliding-window layers use RoPE, while the full-attention layers use NoPE (Cohere et al., 1 Apr 2025).

This architecture is presented as a serving-oriented compromise between performance and efficiency. The model is trained and evaluated up to 256k context, with cooldown stages at 8k for the first 30,000 steps, 32k for 10,000 steps, 128k for 5,000 steps, and 256k for 5,000 steps (Cohere et al., 1 Apr 2025). The systems account emphasizes lower KV-cache memory than full-attention baselines and practical deployment on two A100s or H100s, with throughput up to 156 tokens/sec (Cohere et al., 1 Apr 2025).

For the purposes of interpreting Command A+, the importance of this section is methodological rather than nominal. The available evidence identifies the Command A family, as documented, with long-context enterprise inference, tool use, multilingual operation, and deployment efficiency, not with mixture-of-experts routing, recurrent state-space designs, or low-latency control loops (Cohere et al., 1 Apr 2025).

3. Training regime and model-merging framework

Command A is trained through a decentralized post-training recipe that alternates centralized stages with multiple expert tracks. The documented sequence is: Instruct Model, six SFT Expert Models, SFT Soup Model, six RL Expert Models, RL Soup Model, and a final Polished Model (Cohere et al., 1 Apr 2025). The six expert tracks are Code, Safety, RAG, Math, Multilingual, and General Long-Context (Cohere et al., 1 Apr 2025).

Model merging is a central mechanism in this design. The paper defines merging as

$\theta_{\text{merged} = f(\theta_1, ... ,\theta_K)$

and reports that the practical method used is mostly linear merging or weight averaging, with merge weights chosen by manual search (Cohere et al., 1 Apr 2025). It further states that more complex methods such as SLERP and task vectors did not produce significant improvement over linear merging (Cohere et al., 1 Apr 2025). Merging is used for both cross-capability expert merging and seed merging, and also for capability recovery, especially for long context (Cohere et al., 1 Apr 2025).

The broader training pipeline combines publicly available web text, publicly available code, internally generated synthetic datasets, instruction-tuning datasets obtained from human annotators, and high-quality vendor data (Cohere et al., 1 Apr 2025). It uses an NVIDIA H100 GPU cluster and an internal JAX-based distributed training framework built on GSPMD, with a four-axis mesh over DP, FSDP, SP, and TP (Cohere et al., 1 Apr 2025). The precision recipe stores main weights and optimizer states in FP32, casts model weights to BF16 or FP8 before compute, retains sensitive operations in FP32, and computes attention in BF16; the paper notes that pure FP8 training caused a small but non-trivial degradation in downstream performance, leading to a BF16 warm start (Cohere et al., 1 Apr 2025).

If Command A+ were a later or stronger member of the same family, a plausible implication is that it would inherit this enterprise-first emphasis on capability specialization, post-training decomposition, and merge-based capability consolidation. That remains an inference rather than a documented property of an explicitly named A+ model (Cohere et al., 1 Apr 2025).

4. Enterprise capabilities and benchmark profile

The documented capability profile of Command A is unusually broad for an enterprise-oriented model. The paper reports performance on academic, enterprise, multilingual, tool-use, code, SQL, and long-context evaluations, including MMLU 85.5, MMLU-Pro 69.6, GPQA 50.8, IFEval 90.9, InFoBench 94.9, BFCL Overall 63.8, TauBench 51.7, NTREX 68.8, RepoQA 92.6, Spider test 80.2, Bird 59.5, and MATH 80.0 (Cohere et al., 1 Apr 2025). On the generative enterprise benchmark it reports 94.2%, and on enterprise RAG it reports a Llama Index Correctness average of 4.73 together with Answerable Accuracy 96% and Unanswerable Accuracy 91% (Cohere et al., 1 Apr 2025).

The same paper presents Command A as agent-optimised, multilingual-capable, and designed for retrieval-augmented generation with grounding and tool use. Training examples for tool use include a user prompt, available tools, custom instructions, a reasoning step, tool calls, tool outputs, and a final response with citations to tool outputs (Cohere et al., 1 Apr 2025). Multilingual support is stated for 23 key languages of global business: English, French, Spanish, Italian, German, Portuguese, Japanese, Korean, Arabic, Chinese, Russian, Polish, Turkish, Vietnamese, Dutch, Czech, Indonesian, Ukrainian, Romanian, Greek, Hindi, Hebrew, and Persian (Cohere et al., 1 Apr 2025).

These results matter for the interpretation of Command A+ because they delimit the empirically established identity of the underlying family. The literature does not depict Command A as a narrowly conversational model; rather, it is documented as a grounded, tool-using, long-context, multilingual system optimized for business workflows, enterprise RAG, and controllable deployment (Cohere et al., 1 Apr 2025).

5. Natural-language command and control in physical AI

A separate but highly relevant sense of “command” appears in the drone-control literature. "A Universal LLM -- Drone Command and Control Interface" presents a systems integration architecture that uses MCP as a standard interface layer between LLMs and an operational drone-control backend (Ramos-Silva et al., 21 Jan 2026). The architecture is LLM-agnostic on the top side and MAVLink/MAVSDK-based on the bottom side, with a cloud Ubuntu Linux instance called DroneServer mediating between natural-language instructions and drone actions over HTTP, MAVSDK, MAVLink, and TCP/IP (Ramos-Silva et al., 21 Jan 2026).

The paper exposes 45 tools in total—34 equivalent to MAVSDK-backed functions and 11 custom tools—covering Flight Control, Safety, Navigation, Mission Management, Telemetry, Parameter Management, and Other (Ramos-Silva et al., 21 Jan 2026). It demonstrates flight control of a real unmanned aerial vehicle, extensive SITL testing with ArduPilot SITL on Ubuntu 22.04, Google Maps MCP integration for prompts such as “Fly to the nearest grocery store,” and compatibility with Anthropic Claude, the OpenAI GPT family via OpenAI Agents SDK, Google Gemini, and local/open-source models such as Llama 3.2 / 3.3 70B, Qwen 2.5 72B, Ollama-hosted models, and qwen2.5-7b-instruct via LM Studio (Ramos-Silva et al., 21 Jan 2026).

This work does not document Command A or Command A+ specifically, but it clarifies the operational meaning of command-oriented LLMs in physical systems. The paper shows that natural-language command and control is not reducible to one-shot function calls: naive tool-calling can be unsafe because drones require sequencing, waiting, monitoring, and state-aware execution (Ramos-Silva et al., 21 Jan 2026). It also states that there should always be a human in the loop for possible manual override, that the implementation is not a certified safety interlock system, and that long missions beyond roughly 5–10 minutes were not followed well by current LLM behavior (Ramos-Silva et al., 21 Jan 2026). This suggests that any command-centric model family, including a hypothetical Command A+, must be interpreted not only through benchmark scores but also through the protocol bridges, telemetry loops, and safety semantics required by real command execution.

6. Command telemetry, customization, and user adaptation

Another command-centric line of work concerns the representation and adaptation of user behavior. "SimCURL: Simple Contrastive User Representation Learning from Command Sequences" models longitudinal command streams of software users, segments them into sessions, and learns user embeddings through a user-session network and session dropout (Chu et al., 2022). The reported unlabeled dataset consists of 199,996 users, 1,255,529 sessions, 3,164 unique commands, and 580,028,669 commands; the labeled subset contains 12,612 users, 183,794 sessions, and 58,616,884 commands from Autodesk Fusion 360 telemetry (Chu et al., 2022).

The empirical results show that command traces encode robust signals about user experience and expertise. In the full-label setting, SimCURL reports 37.39 ± 0.41 Accuracy and 35.24 ± 0.68 F1 on 8-class experience classification, 68.67 ± 0.08 Accuracy and 60.26 ± 0.07 F1 on 10 three-class expertise subtasks, and 53.03 ± 0.22 Accuracy and 47.75 ± 0.37 F1 overall (Chu et al., 2022). In the few-shot setting using only 6.25% of labeled training data, it reports 50.26 ± 0.35 Accuracy and 45.04 ± 0.39 F1 overall (Chu et al., 2022). The model’s practical lesson is that command usage contains stable behavioral structure that can be learned from unlabeled logs.

A complementary perspective comes from "An Empirical Investigation of Command-Line Customization," which mines 2,204,199 shell alias definitions, 2,534,167 commands, and 3,630,423 arguments from GitHub (Schröder et al., 2020). The paper derives a taxonomy with three top-level classes—Shortcuts, Modifications, and Scripts—and nine practices including Nicknaming Commands, Abbreviating Subcommands, Bookmarking Locations, Substituting Commands, Overriding Defaults, Colorizing Output, Elevating Privilege, Transforming Data, and Chaining Subcommands (Schröder et al., 2020). It reports, for example, 321,546 Bookmarking Locations aliases, 319,239 Overriding Defaults aliases, 182,623 Colorizing Output aliases, and 93,683 Elevating Privilege aliases (Schröder et al., 2020).

Together, these studies broaden the meaning of command in ways relevant to any Command A+ discussion. Command systems are not only about model-side tool use; they are also about persistent user traits, workflow regularities, contextual defaults, safety preferences, and the compression of repeated friction into reusable command patterns (Chu et al., 2022, Schröder et al., 2020). A plausible implication is that a mature command-centric model family would benefit from all three layers simultaneously: enterprise-grade language modeling, protocol-aware execution, and learned adaptation to real command behavior.

7. Limitations, caveats, and restrained interpretation

The most important limitation is documentary. The available literature does not specify whether Command A+ exists as a distinct model, and it does not provide its size, architecture, tokenizer, language support, context length, or benchmark sheet (Cohere et al., 1 Apr 2025). For that reason, any article on Command A+ must avoid collapsing Command A documentation into direct claims about A+.

The Command A report also states several broader caveats. Academic benchmarks are described as imperfect; human evaluation is noisy; model merging is coarse and does not guarantee a local optimum; code capability preservation under merging is weaker than some other domains; reward models can memorize; and safety trade-offs between answering and refusing remain real (Cohere et al., 1 Apr 2025). The drone-interface paper adds operational caveats: no formal quantitative safety evaluation, no control-theoretic derivation, approximately 5k tokens consumed by 45 tool definitions, and unreliable long-horizon monitoring under current LLM behavior (Ramos-Silva et al., 21 Jan 2026). The user-modeling and shell-customization papers add dataset-level caveats, including lossy session bag-of-commands representations in SimCURL and the fact that public GitHub dotfiles may not represent all shell users (Chu et al., 2022, Schröder et al., 2020).

The most defensible bottom-line interpretation is therefore narrow. Command A+ is not a directly documented arXiv object in the supplied corpus; Command A is the documented reference point, and it is characterized by enterprise deployment, grounded RAG, tool use, multilingual support, long context, and a decentralized merge-heavy training pipeline (Cohere et al., 1 Apr 2025). Adjacent command-oriented research shows that command-capable AI systems extend beyond text generation into MCP-mediated physical control, telemetry-derived user representations, and command-line workflow adaptation (Ramos-Silva et al., 21 Jan 2026, Chu et al., 2022, Schröder et al., 2020). This suggests that the technical significance of any future or externally named Command A+ would depend not only on model-scale improvements, but also on how it engages with the broader command stack: tools, protocols, users, and safety.