Cohere Command A+: the first fully Apache 2.0 enterprise AI model that runs on 2 H100s (May 2026)
Cohere released Command A+ on May 20, 2026—a 218B parameter MoE model (25B active) with native citation generation, W4A4 lossless quantization, and full Apache 2.0 licensing. Runs on a single NVIDIA Blackwell B200 or just 2 H100 GPUs. First fully Apache-licensed frontier model from Cohere, positioning sovereign AI as accessible to enterprises and nations.
CohereCommand A+Open SourceApache 2.0Enterprise AISovereign AI
On May 20, 2026, Cohere released Command A+—a 218 billion parameter Sparse Mixture-of-Experts (MoE) language model with 25 billion active parameters and full Apache 2.0 open-source licensing. The model features native citation generation (explicit grounding spans linking every claim to source documents), W4A4 lossless quantization (enabling deployment on just 2 NVIDIA H100 GPUs), and 48-language support with improved efficiency in non-European languages. Command A+ is Cohere's first fully Apache-licensed frontier model, positioning sovereign AI as accessible to enterprises and nations seeking to control their own AI infrastructure. The release marks a breakthrough in quantization techniques and a strategic shift toward open-weight models for critical infrastructure.
This article is a field guide: what Command A+ is, key features, benchmarks, sovereign AI context, deployment options, and when to choose Command A+ over closed models.
TL;DR
Question
Short answer
What is it?
A 218B parameter MoE model (25B active) with native citations, W4A4 quantization, and full Apache 2.0 license—first fully open frontier model from Cohere.
Announced
May 20, 2026 by Cohere.
Key innovation
W4A4 lossless quantization—4-bit weights + activations with no quality degradation, enabling 2-H100 deployment.
Native citations
Generates explicit grounding spans linking every factual claim to specific source documents or database rows.
Performance
2× faster output speed, 30% lower latency vs previous Command A models. Competitive with GPT-OSS on benchmarks.
Languages
48 world languages with improved efficiency in non-European languages (Arabic, Hindi, Chinese, etc.).
Deployment
Runs on a single NVIDIA Blackwell B200 or 2 H100 GPUs. Available in BF16, FP8, and W4A4 formats.
License
Full Apache 2.0—not just weights, but all components. No restrictions on commercial or sovereign use.
Command A+ is a decoder-only Sparse Mixture-of-Experts (MoE) Transformer with 218 billion total parameters, with only 25 billion active during any given generation step. This design balances frontier performance with efficient inference—generating high-quality outputs while using a fraction of the compute required by dense models.
Core capabilities:
Native citation generation with explicit grounding spans
Complex reasoning and multi-step agentic workflows
48-language support with improved non-European efficiency
RAG-optimized with low-latency retrieval integration
Architecture highlights:
MoE design: 218B total, 25B active (reduces memory bandwidth)
Quantization: Available in BF16 (16-bit), FP8 (8-bit), W4A4 (4-bit)
Context window: TBD (not specified in initial release docs)
Training data: Enterprise-grade corpus (not disclosed)
Key differentiator: Command A+ is the first fully Apache 2.0 licensed frontier model from Cohere—not just model weights, but all components (tokenizer, config, training recipes where applicable). This enables sovereign AI use cases where nations and enterprises need full control without vendor lock-in.
Feature 01: Native citation generation with grounding spans
Problem: Most LLMs hallucinate or provide vague "sources" that don't match the actual content. Post-hoc retrieval systems bolt citations onto generated text, but the mapping is often wrong.
Solution: Command A+ generates explicit "grounding spans" natively during inference—each factual claim is directly linked to the specific source document or database row it pulled the information from.
Example (from Cohere blog):
User query: "What were the revenue figures for Q4 2025?"
Command A+ output: "Q4 2025 revenue was $2.3B, up 15% YoY. [Source: Q4_2025_earnings.pdf, page 3, paragraph 2]"
Grounding span:{"source": "Q4_2025_earnings.pdf", "page": 3, "paragraph": 2, "text": "Revenue for the quarter ending December 31, 2025 totaled $2.3 billion, reflecting a 15% increase compared to Q4 2024."}
Why this matters:
Reduced hallucination (model can't cite sources that don't exist)
Verifiable outputs (users can check the exact source passage)
Trust in critical applications (legal, medical, financial workflows)
How it works:
During training, Command A+ learns to jointly model the generation of text and the retrieval of grounding spans. At inference, the model's attention mechanism explicitly tracks which source tokens contributed to each generated token, producing citations as a byproduct of generation (not a post-hoc step).
Use cases:
Enterprise RAG (retrieval-augmented generation over internal docs)
Legal research (cite case law, statutes, precedent with exact passages)
Medical diagnosis assistants (ground recommendations in clinical guidelines)
Problem: Most quantization techniques (e.g., GPTQ, AWQ) introduce quality degradation—perplexity increases, reasoning degrades, citations become less accurate.
Solution: Cohere's W4A4 quantization compresses both weights (W) and activations (A) to 4 bits while maintaining lossless performance relative to the BF16 baseline.
Technical details:
W4A4: 4-bit weights + 4-bit activations
Compression ratio: ~75% reduction in memory footprint (218B → ~55GB)
Inference speed: 2× faster output, 30% lower latency vs previous Command A models
Quality: No measurable perplexity degradation on internal benchmarks
Why this is a breakthrough:
Most 4-bit quantization schemes (GPTQ, AWQ) only compress weights, leaving activations in FP16/BF16. Command A+ compresses both, enabling deployment on 2 H100 GPUs instead of 8+.
Deployment options:
Format
Precision
VRAM required
Speed
Quality
BF16
16-bit
~400GB (8+ H100s)
Baseline
Baseline
FP8
8-bit
~200GB (4 H100s)
1.5× faster
Minimal loss
W4A4
4-bit
~80GB (2 H100s)
2× faster
Lossless
Single-node deployment:
NVIDIA Blackwell B200: 192GB HBM3e → fits W4A4 in single GPU
2× H100: 80GB each → 160GB total, fits W4A4 comfortably
Why this matters for sovereign AI:
Nations and enterprises can deploy frontier AI on affordable hardware (2 H100s cost ~$60K vs 8 H100s at ~$240K). This lowers the barrier to sovereign AI infrastructure.
Feature 03: 48-language support with non-European efficiency
Problem: Most LLMs are optimized for European languages (English, Spanish, French, German) and underperform on non-European languages (Arabic, Hindi, Chinese, Japanese, etc.).
Solution: Command A+ is trained with native support for 48 world languages and improved tokenization efficiency in non-European scripts.
Key improvements:
Arabic, Hindi, Chinese, Japanese, Korean: 20-30% fewer tokens per sentence vs GPT-4
Code-switching: Better handling of mixed-language text (e.g., English + Hindi in same sentence)
Cultural context: Training data includes regional knowledge graphs, not just translated English text
Example (from Cohere press release):
Input (Hindi): "भारत की राजधानी क्या है?" (What is the capital of India?)
Command A+ output: "भारत की राजधानी नई दिल्ली है। [स्रोत: भारत सरकार की आधिकारिक वेबसाइट]"
Translation: "The capital of India is New Delhi. [Source: Official website of the Government of India]"
Native citation: Grounding span points to Hindi-language government source, not English Wikipedia.
Why this matters:
Sovereign AI for non-Western nations (India, Saudi Arabia, Japan, etc.) can deploy AI in their own languages without relying on English-centric models
Multilingual enterprises (e.g., global banks, UN agencies) can process documents in native languages
Legal document review (search case law, extract clauses, draft summaries)
Feature 05: Full Apache 2.0 licensing—sovereign AI
Problem: Most "open" models release weights only under restrictive licenses (e.g., Llama's "acceptable use policy," Mistral's tiered licensing). This blocks sovereign AI use cases where nations/enterprises need full control.
Solution: Command A+ is released under full Apache 2.0 license—not just weights, but all components (tokenizer, config, training recipes where applicable). No restrictions on commercial, government, or military use.
What this enables:
Sovereign AI infrastructure (nations can run frontier AI on-premises without vendor lock-in)
Critical infrastructure (banks, hospitals, defense agencies can deploy without external dependencies)
Custom fine-tuning (modify the model for domain-specific tasks)
On-premises deployment (no data leaves your network)
Cohere's stated mission (from press release):
"Command A+ advances Cohere's mission to make sovereign AI a technological reality—giving enterprises and nations the power to control their own AI infrastructure."
Comparison to other licenses:
Model
License
Commercial use
Modifications
Redistribution
Sovereign use
Command A+
Apache 2.0
✅ Unlimited
✅ Yes
✅ Yes
✅ Yes
Llama 3
Custom (Meta)
✅ With restrictions
✅ Yes
❌ Restricted
❌ Restricted
Mistral Large
Mistral AI License
✅ Tiered
✅ Limited
❌ No
❌ No
GPT-4
Closed
❌ API only
❌ No
❌ No
❌ No
Gemini
Closed
❌ API only
❌ No
❌ No
❌ No
Why this matters:
For the first time, a frontier-class model (218B params, competitive with GPT-4-class) is available with zero licensing friction for sovereign AI use.
Benchmarks and performance
From Cohere blog:
Speed and efficiency
2× faster output speed vs previous Command A models
30% lower latency (time to first token + generation)
Runs on 2 H100s (W4A4 quantization)
Reasoning and agentic tasks
Across-the-board improvements for agentic, reasoning, and multi-step tasks
Competitive with GPT-OSS (unclear which variant—GPT-4? GPT-5?)
Tool-calling accuracy: Comparable to GPT-4-class models
Multimodal document processing
Native handling of structured data (tables, CSVs, JSON)
Grounding spans link claims to specific cells/rows
Better than previous Command A on document QA benchmarks
Missing public benchmarks:
As of May 22, 2026, Cohere has not released:
MMLU, HellaSwag, TruthfulQA scores
Head-to-head comparison vs Llama 3, Mistral Large, GPT-4
Citation accuracy metrics (precision/recall on grounding spans)
Community reactions (from X/Twitter):
Positive: "Running on 2 H100s is huge for practical deployment."
Skeptical: "Where are the benches vs SOTA open models (Qwen series)?"
Excited: "If this is better than Gemini 3.1 Flash Lite, that's game-changing for fast agent products."
Use cases: sovereign AI, enterprise RAG, critical infrastructure
01. Sovereign AI for nations
Problem: Countries like India, Saudi Arabia, Japan, and EU nations want to deploy frontier AI without dependency on US cloud providers (AWS, Azure, GCP) or closed APIs (OpenAI, Anthropic, Google).
48-language support (native Hindi, Arabic, Japanese, etc.)
Native citations (verifiable outputs for government/legal use)
Example: India's National AI Infrastructure can deploy Command A+ on local data centers, process Hindi/Tamil/Telugu documents, and generate outputs with citations to Indian legal precedent—all without data leaving the country.
02. Enterprise RAG over internal documents
Problem: Enterprises have terabytes of internal docs (contracts, emails, Confluence, Slack) but can't send them to OpenAI/Anthropic due to confidentiality.
Solution: Command A+ runs on-premises with native citations, enabling:
Secure RAG over sensitive documents
Grounding spans linking answers to exact source passages
Low-latency inference (2× faster output vs previous Command A)
Example: A law firm deploys Command A+ on-premises, indexes 50K case files, and lets associates query "What are the precedents for X?" with answers citing specific case numbers and paragraphs.
Problem: Defense agencies, hospitals, and banks can't use cloud APIs for classified/sensitive workflows.
Solution: Command A+ runs air-gapped on local hardware:
No internet connection required (model weights stored locally)
Full control over data (no external API calls)
Native citations for audit trails
Example: A hospital deploys Command A+ for clinical decision support—physicians query "What are the treatment guidelines for X?" and get answers citing exact NIH/CDC guidelines with page numbers.
Deployment: 2 H100s, vLLM, transformers
Hardware requirements:
Format
GPUs
VRAM
Throughput
BF16
8× H100
~400GB
Baseline
FP8
4× H100
~200GB
1.5× faster
W4A4
2× H100
~80GB
2× faster
Blackwell B200
1× GPU
192GB
2× faster
Software stack:
vLLM: Recommended for production serving (high throughput, low latency)
transformers: Standard HF API works (slower, good for prototyping)
LiteLLM: Unified API for multiple providers
Installation:
bash
# W4A4 quantized model
pip install transformers torch
from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained(
"CohereLabs/command-a-plus-05-2026-w4a4",
device_map="auto",
torch_dtype="auto",
)
tokenizer = AutoTokenizer.from_pretrained("CohereLabs/command-a-plus-05-2026-w4a4")
inputs = tokenizer("What were Q4 2025 revenues?", return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=512)
print(tokenizer.decode(outputs[0]))
Model capabilities, benchmark scores, and deployment options may change with future releases. Treat this as May 22, 2026 context—verify performance claims and license terms at cohere.com before production deployment. Command A+ is fully Apache 2.0 licensed; commercial, government, and military use is permitted without restriction.
