Cohere's Command A+ Runs a 218B Open Model on Just Two GPUs
Cohere open-sources Command A+, a 218B MoE model that runs on just two H100s with near-lossless 4-bit quantization and built-in citations
- Command A+ released open-source: Cohere's 218B sparse MoE model is now available under Apache 2.0 with no commercial restrictions. Download on Hugging Face
- Runs on 2x H100s: W4A4 quantization with near-lossless quality via Quantization-Aware Distillation (QAD) keeps attention layers at full precision
- Consolidates 5 models into 1: Replaces Command A, Command A Reasoning, Command A Vision, and Command A Translate in a single deployment
- Massive agentic gains: τ²-Bench Telecom jumped from 37% to 85%; Terminal-Bench Hard from 3% to 25% vs. Command A Reasoning
- Native citations built in: Every factual claim is linked to its source document via grounding span tags, critical for regulated-industry RAG
- API pricing: $2.50/$10 per M tokens; also available via Cohere's Model Vault for private managed inference
Cohere just dropped what may be the most practically deployable open-weight frontier model yet. Command A+ is a 218-billion-parameter sparse Mixture-of-Experts (MoE) model released under the Apache 2.0 license, meaning full commercial use with zero revenue caps or usage restrictions. The headline number that matters most: it runs on as few as two NVIDIA H100 GPUs at 4-bit quantization, with virtually no quality loss.
One model to replace five
Born from a year of deploying North with enterprise customers, Command A+ surpasses every previous generation in the Command series and unifies their capabilities into a single scalable model. That consolidation is a bigger deal than it sounds. Previously, Cohere maintained separate specialized variants:
- Command A , base model, tool use, 23 languages
- Command A Reasoning , extended thinking, chain-of-thought
- Command A Vision , image understanding
- Command A Translate , multilingual translation
Command A+ consolidates capabilities that previously lived across separate Command A, Command A Reasoning, Command A Vision, and Command A Translate variants into a single model. For teams managing multiple fine-tuned or specialized deployments, that is a significant operational simplification.
The architecture: sparse by design
Command A+ uses a sparse Mixture-of-Experts decoder with 218 billion total parameters organized across 128 experts. Each token activates 8 experts plus 1 shared expert, meaning 25 billion parameters are active per inference call, not 218 billion. This is the core trick behind its hardware efficiency: the model is large in total capacity but cheap to run token-by-token.
It requires far less compute to run than proprietary giants like OpenAI's GPT-5.5 and Anthropic's Claude Opus 4.7, which are estimated by third-party observers to be in the trillions of parameters. The MoE approach routes each token only to the expert subnetworks best suited to handle it, leaving the rest dormant.
The quantization breakthrough
Getting a 218B model to fit on two H100s without degrading quality is the real engineering story here. The genuinely innovative piece is Quantization-Aware Distillation (QAD): rather than applying uniform quantization across the model, QAD preserves attention pathway weights at full precision while quantizing only the MoE expert layers. The result is near-lossless W4A4 quantization, with the compressed model retaining quality while running on significantly less hardware.
QAD works by training a quantized "student" model to match the output distribution of the full-precision "teacher" model, using fake quantization operators during the forward pass. The intuition: reasoning models are especially sensitive to quantization errors because long decoding chains compound small per-token mistakes. By keeping attention at full precision and only compressing the expert layers (which dominate parameter count but are less error-sensitive), Cohere gets the memory savings without the quality hit.
Three quantization tiers are available on Hugging Face:
- BF16 (16-bit): 4x B200 or 8x H100
- FP8 (8-bit): 2x B200 or 4x H100
- W4A4 (4-bit): 1x B200 or 2x H100 , recommended for most deployments
Speed that compounds
The efficiency gains stack across multiple techniques. At the same quantization and concurrency levels, Command A+ delivers up to 63% higher output tokens per second and reduces time to first token by up to 17% compared with Command A Reasoning. The W4A4 quantization contributes an additional 47% increase in speed and a 13% reduction in latency.
Speculative decoding, optimized specifically for the MoE architecture, delivers an additional 1.5 to 1.6x inference speedup for both text and multimodal inputs. Speculative decoding works by using a smaller "draft" model to predict several tokens ahead, then verifying them in parallel with the main model. MoE architectures benefit from this more than dense models because their compute profile has a sweet spot at moderate batch sizes where the speedup is non-monotonic.
Command A+ also ships with a new tokenizer that compresses non-European languages more efficiently. Tokenization efficiency improved by 20% for Arabic, 16% for Korean, and 18% for Japanese compared to the previous tokenizer. Fewer tokens per response means lower inference cost at scale.
What it's actually good at
The benchmark improvements over Command A Reasoning are dramatic in agentic settings:
- Scores jumped from 37% to 85% on the agent benchmark τ²-Bench Telecom and from 3% to 25% on Terminal-Bench Hard
- Agentic Question Answering accuracy improved by 20% and spreadsheet analysis quality by 32% over Command A Reasoning on internal North evaluations
- MathVista scores improved from 73.5% to 80.6%, and CharXiv reasoning improved from 46.9% to 52.7%
- Multilingual coverage expanded from 23 to 48 languages, with gains in machine translation and multilingual reasoning
Command A+ scored 37 on the Artificial Analysis Intelligence Index, outperforming other leading open models. On that index, it lands roughly on par with Claude 4.5 Haiku, Gemma 4 31B, and Mistral Medium 3.5. It is not competing with frontier closed models on raw capability, but that is not the point.
The enterprise feature that matters most
Command A+ features native citation generation. When the model retrieves information from an external tool, it generates explicit grounding spans, using special tags embedded in the output to directly link every factual claim to the specific source document or database row it pulled the information from.
In practice, the output looks like this:
On 29th September 2023, total sales were <co>10000</co: 0:[0]>
and units sold were <co>250.</co: 0:[0]>For enterprises in heavily regulated industries like finance, healthcare, or legal, this traceability is the difference between an interesting prototype and a production-ready application. Grounded citations are table stakes for compliance-sensitive RAG pipelines, and having them baked into the model rather than bolted on post-hoc is a meaningful architectural advantage.
Where it fits in the market
The sovereign AI angle is Cohere's sharpest differentiator. Command A+ requires far less compute than proprietary models from OpenAI and Anthropic, which are estimated to be in the trillions of parameters. For governments, financial institutions, and healthcare systems that cannot send data to a third-party API, a model that runs entirely within their own infrastructure at two H100s is a genuinely different proposition.
Command A+ is the first fully Apache 2.0 licensed frontier model from Cohere, covering not just model weights but all components including the tokenizer and config. That licensing clarity matters for legal and procurement teams at large organizations, who have historically been cautious about open models with ambiguous commercial terms.
How to get started
The model is available now through several paths:
- Hugging Face: Download weights directly at CohereLabs/command-a-plus-05-2026-w4a4 (W4A4 recommended)
- Cohere API: Available via the Cohere dashboard with a free API key. API pricing is $2.50 per million input tokens and $10.00 per million output tokens.
- Model Vault: Cohere's managed, dedicated inference environment for enterprise deployments
- HF Space: Try it without downloading weights at the hosted Hugging Face Space
For vLLM users, the W4A4 variant requires vllm>=0.21.0 and Cohere's cohere_melody library for accurate response parsing. The model supports both vLLM and Transformers out of the box, with tool use handled through standard chat templates.
pip install vllm>=0.21.0 cohere_melody>=0.9.0
vllm serve CohereLabs/command-a-plus-05-2026-w4a4 \
-tp 2 \
--tool-call-parser cohere_command4 \
--reasoning-parser cohere_command4 \
--enable-auto-tool-choiceThe model's 128K input context and 64K output context make it viable for long-document RAG, multi-turn agentic loops, and large codebase analysis. For teams building private AI infrastructure who previously had to choose between capability and hardware cost, Command A+ meaningfully shifts that tradeoff.