Model

Command A2

OPEN

Cohere · Command A family · released Sep 16, 2026

Cohere's September 16, 2026 successor to May's Command A+ — 340B/32B MoE under Apache 2.0, keeping the line's enterprise-multilingual positioning (48 languages) and, like A+, shipped weights-first with no first-party metered price at launch (price: null, same as A+). The launch table is a straight +3-8 point lift across the A+ suite with first-time SWE-Bench Pro and DeepSearchQA rows; Terminal-Bench 2.1 remains the family's weak spot at 29.4.

ReasoningCodingVisionFunction callingTool useAgentic

1885.6

Elo · rank #95

Parameters

340B

Active params

32B (MoE)

Context

256K tokens

Architecture

340B-total/32B-active sparse MoE, Apache 2.0 open weights; 256K-token context, 48-language coverage

License

Apache 2.0

Languages

48+

API price (in/out)

No hosted API

Modalities

text · vision

Benchmark results

Bar shows position within the tracked field; marker = field best

AIMEMath93.2%#31

best: GPT-5.2 · 100.0%

CharXivVision61.4%#39

best: Qwen3.8-Flash-Next · 90.6%

DeepSearchQAAgents71.2%#6

best: GPT-5.6 Sol · 93.0%

GPQA DiamondReasoning81.7%#87

best: GPT-6 Astra · 96.0%

Humanity's Last ExamReasoning16.8%#87

best: Claude Opus 5 · 64.7%

IFBenchReasoning79.4%#13

best: Inkling-Medium · 85.3%

LiveCodeBenchCoding83.7%#36

best: DeepSeek-V4.1 (Think Max) · 94.2%

MathVistaVision84.9%#11

best: Seed 2.2 Pro · 91.4%

MMLU-ProKnowledge82.6%#48

best: Claude Fable 5 · 91.5%

MMMU-ProVision69.8%#43

best: Claude Opus 4.7 · 85.5%

MMMUVision79.6%#26

best: Claude Fable 5 · 89.3%

SWE-bench ProCoding46.8%#64

best: Claude Fable 5.1 · 81.2%

τ²-Bench TelecomAgents88.4%#22

best: Claude Opus 4.6 · 99.3%

Terminal-Bench 2.0Coding29.4%#83

best: Gemini 3.8 Flash · 89.4%

Run it locally

VRAM @ Q4

—

VRAM @ FP16

—

Fits on (Q4)

Multi-node cluster required

Throughput data unavailable.

Quantizations

—

Fine-tune it

Permissive

QLoRA231.2 GB2× H200 141GB

LoRA724.2 GB4× B200 192GB

Full fine-tune5457.0 GBbeyond 8× B200

QLoRA SFT on ~10k samples ≈ $16.55 (2× H200 141GB)

Plan a fine-tune →

Command A family

Elo progression across releases

Command AMar 20251083.0 Command A+May 20261692.8 Command A2Sep 20261885.6

Resources

HFCohereLabs/command-a2 DOCShttps://docs.cohere.com/docs/command-a2 APIAvailable via Cohere API

API price weights · each benchmark row carries its own source badge (see methodology)