SetupAI
繁
SourceStalecollected in 15m

Cohere releases Apache 2.0 open-weight model Command A+

Read original on VentureBeat
#moe#quantization#sovereign-ai#llm

First Apache 2.0 frontier-grade MoE model optimized for local deployment and complex reasoning.

30-Second TL;DR

What Changed

218B parameter sparse MoE architecture with only 25B active parameters per step.

Why It Matters

This release enables enterprises to run frontier-grade reasoning models on-premises, reducing dependency on proprietary cloud APIs. It sets a new standard for open-source models that prioritize hardware efficiency and sovereign data control.

What To Do Next

Download the Command A+ weights from Hugging Face and benchmark the W4A4 quantization performance against your current local inference setup.

Who should care:Developers & AI Engineers

Key Points

  • •218B parameter sparse MoE architecture with only 25B active parameters per step.
  • •Released under a permissive Apache 2.0 license on Hugging Face for sovereign AI deployment.
  • •Supports advanced quantization formats including W4A4 to minimize memory footprint without sacrificing reasoning quality.
Key numbers110%30%

Deep Insight

Background and context from public sources — not the original article. 22 sources cited.

Enhanced Key Takeaways

  • •Command A+ is the final model in the Command A family and Cohere's inaugural Mixture-of-Experts (MoE) model, integrating vision inputs alongside text for multimodal capabilities.
  • •The model significantly expands multilingual support to 48 languages, encompassing all official EU languages, which more than doubles the language coverage of previous Cohere models.
  • •Optimized for production-grade deployments, Command A+ can be deployed efficiently on minimal hardware, requiring as few as one B200 or two H100 GPUs, and demonstrates up to a 110% increase in throughput and a 30% decrease in latency compared to Command A Reasoning.
  • •Its Apache 2.0 license is specifically designed to facilitate sovereign AI deployments, offering governments and regulated industries full transparency, control, and the ability to deploy on-premises or in private clouds without external data transmission.
  • •Command A+ is noted for its strong performance in pure reasoning and mathematics, competing with larger models like DeepSeek V4 Pro, although it currently trails some Chinese open-source rivals in deep agentic coding and general intelligence indexing.

Competitor Analysis

Total Parameters
Cohere Command A+
218 Billion
Qwen3.6-35B-A3B
35 Billion
GPT-OSS-120B
117 Billion
Google Gemini 3 Pro (Proprietary)
Sparse MoE (details not specified)
Active Parameters
Cohere Command A+
25 Billion
Qwen3.6-35B-A3B
3 Billion
GPT-OSS-120B
5.1 Billion
Google Gemini 3 Pro (Proprietary)
Sparse MoE (details not specified)
License
Cohere Command A+
Apache 2.0
Qwen3.6-35B-A3B
Open Weights (permissive)
GPT-OSS-120B
Apache 2.0
Google Gemini 3 Pro (Proprietary)
Proprietary
Key Optimizations/Features
Cohere Command A+
Multimodal, Agentic, Reasoning, Multilingual (48 langs), W4A4 Quantization
Qwen3.6-35B-A3B
Agentic Coding, Multimodal Perception, Reasoning
GPT-OSS-120B
Reasoning, Consistent Structured Responses, MXFP4 Quantization
Google Gemini 3 Pro (Proprietary)
Multimodal (text, image, audio, video), Agentic, Long Context Reasoning
Hardware Requirements (Min.)
Cohere Command A+
1 x B200 or 2 x H100s (W4A4)
Qwen3.6-35B-A3B
Not specified, but efficient
GPT-OSS-120B
1 x H100 or AMD MI300X 80GB GPU
Google Gemini 3 Pro (Proprietary)
Runs on Google TPUs
Context Length (Input)
Cohere Command A+
128K tokens
Qwen3.6-35B-A3B
Not specified
GPT-OSS-120B
Not specified
Google Gemini 3 Pro (Proprietary)
1 Million tokens
Output Length (Max)
Cohere Command A+
64K tokens
Qwen3.6-35B-A3B
Not specified
GPT-OSS-120B
Not specified
Google Gemini 3 Pro (Proprietary)
64K tokens
Benchmarks
Cohere Command A+
Competes with DeepSeek V4 Pro on math; trails some Chinese rivals in agentic coding/general intelligence
Qwen3.6-35B-A3B
Outperforms Qwen3.5-35B-A3B, rivals larger dense models
GPT-OSS-120B
Strong reasoning capability
Google Gemini 3 Pro (Proprietary)
Improves over Gemini 2.5 Pro, competitive with GPT 5.1 and Claude Sonnet 4.5 on reasoning, multimodal, coding benchmarks

Technical Deep Dive

  • Architecture: Decoder-only Sparse Mixture-of-Experts (MoE) Transformer Model.
  • Expert Configuration: Comprises 128 experts, with 8 experts actively engaged per token, and includes a single shared expert applied to all tokens.
  • Attention Mechanism: Features interleaved sliding-window attention layers with Rotational Positional Embeddings and global attention layers without positional embeddings, arranged in a 3:1 ratio, a design inherited from Command A.
  • MoE Training: Utilizes a fully dropless training approach and a token-choice router for expert selection.
  • Quantization Support: Offers support for BF16, FP8, and W4A4 quantization formats, with W4A4 being the recommended choice for its superior speed, latency, and reduced hardware footprint without significant quality degradation.
  • Context Window: Supports an input context length of 128,000 tokens and a maximum output length of 64,000 tokens.
  • Multimodal Capability: Integrates native support for vision inputs, allowing the model to process both text and images.
  • Deployment Efficiency: Designed for high-performance inference on as few as one NVIDIA B200 or two NVIDIA H100 GPUs when utilizing W4A4 quantization.

Future ImplicationsAI analysis grounded in cited sources

Command A+ will accelerate the adoption of sovereign AI solutions in regulated industries.
Its Apache 2.0 license, explicit focus on data sovereignty, and efficient deployment options directly address critical concerns for governments and enterprises handling sensitive data, fostering trust and control.
The model's multimodal and advanced agentic capabilities will drive innovation in complex enterprise automation.
By combining vision inputs with robust reasoning and tool-use, Command A+ enables the development of more sophisticated and autonomous AI agents capable of handling intricate business workflows.
Cohere's strategy of releasing powerful open-weight MoE models will intensify competition in the enterprise AI market.
Offering a high-performance, efficient, and openly licensed MoE model challenges proprietary offerings and encourages other vendors to either follow suit with more open models or further differentiate their closed-source solutions.

Timeline

2017
Aidan Gomez co-authors the seminal 'Attention Is All You Need' Transformer paper.
2019
Cohere is founded in Toronto by Aidan Gomez, Nick Frosst, and Ivan Zhang.
2021
Secures Series A funding ($40M) and launches its first commercial API; partners with Google Cloud for TPU infrastructure.
2023
Raises $270M in Series C funding and announces a strategic partnership with Oracle.
2024
Releases Command R+ model and raises $450M in Series D funding, valuing the company at $5.5B.
2025-03
Launches Command A, an enterprise AI model optimized for speed, efficiency, and agentic/multilingual tasks.
2025-09
Introduces Model Vault, a platform for secure enterprise model deployment in isolated environments.
2026-02
Cohere Labs releases the Tiny Aya family of open-weight multilingual models.
2026-04
Agrees to acquire German AI firm Aleph Alpha.
2026-05-20
Cohere releases Command A+, a 218-billion-parameter sparse MoE model under an Apache 2.0 license.
Read original article on VentureBeat

Weekly AI Recap

Read this week's curated digest of top AI events →

PostLinkedIn

AI-curated news aggregator. All content rights belong to original publishers.
Original source: VentureBeat ↗

This is a summary, not the original. Read the source, or get the weekly briefing.

The weekly digest

One email a week. Unsubscribe anytime.