Cohere Ships Command A+ as the Enterprise Open-Model Race Heats Up

# Cohere Ships Command A+ as the Enterprise Open-Model Race Heats Up

As the final week of July 2026 delivered what many researchers are calling the biggest open-model drop in the industry's short history, Cohere has planted a distinctly enterprise flag in the ground. Amid a deluge of Chinese frontier releases — Moonshot's 2.8-trillion-parameter Kimi K3, DeepSeek V4 Pro, and Zhipu's GLM-5.2 — the Toronto- and San Francisco-based lab is pressing its case that raw benchmark supremacy is not what regulated buyers are shopping for. Its answer is Command A+, the company's most powerful model to date and its first release under a full, OSI-approved Apache 2.0 license.

Command A+ is a decoder-only sparse Mixture-of-Experts transformer with 218 billion total parameters, of which only about 25 billion activate on any given generation step. That efficiency is the pitch: through a lossless W4A4 4-bit quantization variant, the model runs on as little as a single B200 or two H100 GPUs, while BF16 and FP8 variants scale up to 8×H100 configurations. It ships with a 128,000-token context window, up to 64,000 output tokens, image input, and support across 48 languages — deliberately modest numbers next to DeepSeek V4 Pro's advertised million-token context, but sized for the on-prem racks Cohere's customers actually own.

"Our first fully open source Apache 2 model :)," co-founder and CEO Aidan Gomez wrote in a short post announcing the release. The understatement belies a strategic shift. Cohere, founded in 2019 by Gomez — a co-author of the original "Attention Is All You Need" transformer paper — Ivan Zhang, and Nick Frosst, has historically kept its weights close. Putting Command A+ under Apache 2.0 lets any organization, from a solo developer to a Fortune 500 bank, deploy, modify, and commercialize the model with no licensing fees and no non-compete clauses.

On benchmarks, Command A+ posts 75.1% on MMMU and 63% on the harder MMMU Pro, 80.6% on MathVista, and 52.7% on CharXiv reasoning. On agentic evaluations it records 85 on Tau2 Telecom and 25 on Terminal-Bench Hard. Cohere says the model delivers up to 63% higher output tokens per second and cuts time-to-first-token by as much as 17% versus its earlier Command A Reasoning model at matched quantization and concurrency. Those figures trail the eye-watering scores from Kimi K3 — which claimed a record 93.5% on GPQA Diamond — but Cohere is not competing for the leaderboard crown.

Why it matters

The story of late July 2026 is really two stories running in parallel. One is a Chinese-led sprint for the open-weight frontier, where MIT-licensed giants keep resetting the ceiling on capability and driving inference costs toward the floor — GLM-5.2 undercuts K3's output price by more than two-thirds, and DeepSeek V4 Pro offers million-token context for under a dollar per million output tokens. The other is a quieter contest for the enterprise, where the buying criteria are governance, provenance, and control rather than the top line of a benchmark table.

Cohere is betting its future on the second contest. Command A+ is engineered around features that regulated industries and governments care about: native citations, in which the model emits explicit "grounding spans" that tie each factual claim back to the specific source document or database row it drew from, and full data sovereignty for deployments that never leave a customer's own infrastructure. Cohere describes the model as "purpose-built for sovereign critical infrastructure," language aimed squarely at banks, insurers, health systems, and public-sector agencies that cannot ship sensitive data to a third-party API.

That niche is defensible precisely because it is unglamorous. Llama's open releases and Mistral's European enterprise push both crowd the same territory, but Cohere's decade-long focus on retrieval-augmented generation, reranking, and grounded output gives it a coherent product story that a headline GPQA score cannot buy. In a week when the marginal capability of open models is arguably commoditizing faster than anyone expected, the durable differentiator may be trust: an auditable, on-prem, Apache-licensed model that a compliance officer can sign off on. The Chinese labs are winning the capability race in public. Cohere is trying to win the procurement race in private.

What to watch

The open question is whether "good enough plus governance" holds up as the capability gap widens. If Kimi K3, DeepSeek V4, and GLM-5.2 continue to pull ahead on reasoning and agentic tasks — and their weights are freely downloadable — enterprises may simply wrap those models in their own retrieval and citation layers rather than pay for Cohere's. Watch three things over the coming quarter: whether Command A+ shows up in named government and financial-sector deployments; whether Cohere publishes transparent enterprise pricing to sit alongside Command A's $2.50-in / $10-out per-million rate; and whether rivals fast-follow with their own native-citation and sovereignty tooling. Cohere has staked out the enterprise open-model lane at exactly the moment it became the most crowded corner of the market. The next few months will reveal whether that lane is a moat or merely a temporary head start.