Cohere Command A+ vs Faraday Machines
At launch, Cohere Command A+ scored 37 on the Artificial Analysis Intelligence Index (pre-v4.3). Every model Faraday Machines shipped in May 2026 — Kimi K2.6 (54), DeepSeek V4 Flash (47), Qwen 3.6 Plus (50), GLM-5.1 (51) — outperformed it by 10 to 17 points on that index. Here are the benchmarks, and the current lineup.
Key takeaways
- Cohere Command A+ scored 37 on the Artificial Analysis Intelligence Index at its May 2026 launch (pre-v4.3), 10 to 17 points behind the open models Faraday Machines shipped then.
- On the rescaled Artificial Analysis Intelligence Index v4.3.2, Command A+ is listed at an estimated 13, pending independent evaluation, while GLM-5.3-Flash scores 42 and MiMo-V2.6-Flash 38.
- Cohere’s changelog calls Command A+, a 218B-parameter (25B active) Apache 2.0 model, the last in the Command A family, and no successor had shipped as of October 2026.
- Command A+ has real strengths, including native citation grounding and 48-language support, but at launch it scored 25% on Terminal-Bench Hard and 86% on AA-Omniscience non-hallucination.
- Faraday Machines ships DeepSeek V4.1 Flash, GLM-5.3-Flash, Qwen3.8-27B, MiMo-V2.6-Flash and MiniMax M3 on-premises from $299 US a month, with Kimi K3 and GLM-5.3 on larger setups.
October 2026 update
This comparison was published on May 29, 2026. Every score in the May tables and stats below comes from the Artificial Analysis Intelligence Index as it stood in May–June 2026 (pre-v4.3), and the green-row models are the ones Faraday Machines shipped then. Since publication:
- Artificial Analysis rescaled its index with v4.3 (announced September 7, 2026). Every model's score dropped, so May scores can't be compared with current ones.
- On v4.3.2, Artificial Analysis lists Command A+ at 13, marked as an estimate pending independent evaluation. Cohere's changelog calls Command A+ the last model in the Command A family, and no successor has shipped.
- Faraday Machines now ships DeepSeek V4.1 Flash, GLM-5.3-Flash (42 on v4.3), Qwen3.8-27B, MiMo-V2.6-Flash (38) and MiniMax M3. Kimi K3 (44) and GLM-5.3 (45) are available on larger setups. See the current lineup below.
- We corrected DeepSeek V4.1 Flash’s size on October 19: it is about 400–445 GB at 4-bit, not ~300 GB, so it runs on Scale as a 1- to 3-bit build and needs a larger setup at 4-bit.
At a Glance: May 2026 (Pre-v4.3 Index)
| Model | Type | Intelligence Index (pre-v4.3) | Context | Params (Active) | License | Best For |
|---|---|---|---|---|---|---|
| Cohere Command A+ | Open | 37 | 128K in / 64K out | 218B (25B active) | Apache 2.0 | Enterprise RAG, citations |
| DeepSeek V4 Flash | Open | 47 | 1M | 284B (13B active) | MIT | Cost-efficient inference, 1M context |
| Qwen 3.6 Plus | Open | 50 | 1M | Hybrid MoE | Apache 2.0 | Tool calling, MCP integration |
| GLM-5.1 | Open | 51 | 200K | 744B (~40B active) | MIT | Regulated industries, multilingual |
| Kimi K2.6 | Open | 54 | 256K | 1T (~32B active) | Modified MIT | Agentic workflows, complex coding |
Intelligence Index scores from Artificial Analysis v4.0 (May 2026), the pre-v4.3 index; not comparable with v4.3 scores. Grey rows = Cohere Command A+. Green rows = Models Faraday Machines shipped in May 2026.
The Current Lineup (October 2026, Index v4.3.x)
| Model | Type | Intelligence Index v4.3.x | Context | Params (Active) | Runs On | Best For |
|---|---|---|---|---|---|---|
| Cohere Command A+ | Open | 13 (AA estimate) | 128K in / 64K out | 218B (25B active) | — | Enterprise RAG, citations |
| DeepSeek V4.1 Flash | Open | — | 1M | 552B (16B active), plus ~196B Engram memory | Scale (512 GB) at 1–3-bit; larger setup at 4-bit | Agentic coding, long-context work |
| GLM-5.3-Flash | Open | 42 | 1M | 320B (18B active) | Scale (512 GB); Growth (256 GB) at 1–2-bit | Business intelligence, documents, multimodal input |
| Qwen3.8-27B | Open | — | 262K | 27B dense | Starter (128 GB) | Fast coding, multilingual, image & video |
| MiMo-V2.6-Flash | Open | 38 | — | 309B (15B active) | Scale (512 GB) | Cost-efficient generalist |
| MiniMax M3 | Open | — | Up to 1M | ~428B (~23B active) | Scale (512 GB), a tighter fit | Long-context, multimodal agents |
| GLM-5.3 | Open | 45 | 1M | ~753B (~40B active) | Scale (512 GB) at 1–2-bit; larger setup at 4-bit | Complex coding, agentic work |
| Kimi K3 | Open | 44 | 1M | 2.8T (104B active) | Larger setups (at 1-bit or 4-bit) | Agentic coding, browsing, frontend |
Intelligence Index scores from Artificial Analysis v4.3.x (v4.3.2 page, October 2026); not comparable with the May 2026 table above. Command A+'s 13 is Artificial Analysis's estimate, pending independent evaluation, so treat any gap to it with caution. — = no v4.3 score (or context figure) in our sources. “Runs On” is the smallest Faraday Machines configuration that holds the model at 4-bit quantization (roughly), or at the lower bit width shown. The lower-bit options are 1- to 3-bit builds, mostly Unsloth’s; they trade some quality, most noticeably on long or precise work.
On today's index, the best open-weight model is MiMo-V2.6-Pro at 46 (available on larger setups). Of the models Faraday ships on its standard tiers, GLM-5.3-Flash scores 42 and MiMo-V2.6-Flash 38. GLM-5.3-Flash's 42 matches GPT-5.6 Terra on Artificial Analysis's v4.3 index at $0.25 per task versus $1.40. Command A+, meanwhile, is still Cohere's latest model.
The Intelligence Gap
Cohere Command A+ is a capable model — especially for its 25B active-parameter footprint. Its native citation grounding, 48-language support, and efficient W4A4 quantization are genuine engineering achievements. But capability is relative, and on the Artificial Analysis Intelligence Index as it stood in May 2026 (pre-v4.3), the gap was stark:
Seventeen points on the Intelligence Index isn't a marginal difference. It's the gap between a model that can handle routine tasks and models that can reason through complex, multi-step problems — the kind your business actually needs AI for. Those gaps were measured on the pre-v4.3 index and don't carry over to today's scale; for the current picture, see the October 2026 lineup above.
Software Engineering: The Benchmark That Matters Most
If you're evaluating AI for coding, developer tooling, or technical automation, SWE-bench Verified was the gold standard at publication. It measures whether an AI agent can resolve real GitHub issues end-to-end. At launch, Cohere Command A+ had no published SWE-bench score. The models Faraday Machines shipped in May 2026 all did:
Cohere Command A+'s 25% on Terminal-Bench Hard (agentic coding) isn't in the same league as Kimi K2.6's 66.7% or DeepSeek V4 Pro's 93.5% on LiveCodeBench. If your team uses AI for development work, Cohere Command A+ will produce meaningfully worse results. Faraday's current coding pick, DeepSeek V4.1 Flash, scores 77.0 on DeepSWE v1.1 and 42.3 on Terminal-Bench 4.0 in its October 17 update (DeepSeek-reported).
Hallucination and Accuracy
For business applications, a model that confidently gives wrong answers is worse than one that says "I don't know." Cohere Command A+ achieves 86% on the AA-Omniscience Non-Hallucination benchmark — decent, but behind the models Faraday Machines shipped at publication:
DeepSeek V4 Pro: 94% Non-Hallucination
8 points ahead of Cohere Command A+. When accuracy matters for legal, financial, or medical use cases, this gap compounds with every query.
DeepSeek V4 Flash: 96% Non-Hallucination
10 points ahead. The budget-oriented Flash model is actually more reliable than Cohere Command A+ on factual accuracy.
Cohere Command A+: 86% Non-Hallucination
14% of factual claims may be incorrect. For RAG use cases — Cohere's stated strength — this means one in seven cited facts could be wrong.
Missing SWE-bench Scores
At launch, Cohere Command A+ had no published SWE-bench Verified or LiveCodeBench scores. When a model doesn't publish on the industry-standard coding benchmark, it's reasonable to assume the results don't flatter it.
Architecture Comparison (May 2026 Lineup)
| Cohere Command A+ | DeepSeek V4 Flash | Qwen 3.6 Plus | GLM-5.1 | Kimi K2.6 | |
|---|---|---|---|---|---|
| Intelligence Index (pre-v4.3) | 37 | 47 | 50 | 51 | 54 |
| Total Parameters | 218B | 284B | Hybrid MoE | 744B | ~1T |
| Active Parameters | 25B | 13B | Hybrid MoE | ~40B | ~32B |
| Context Window | 128K / 64K | 1M | 1M | 200K | 256K |
| License | Apache 2.0 | MIT | Apache 2.0 | MIT | Modified MIT |
| Min Hardware (quantized) | 2× H100 | 2× H100 | 2× H100 | 8× H100 | 4× H100 |
| On Faraday Machines (May 2026) | — | ✓ | ✓ | ✓ | ✓ |
What Cohere Command A+ Gets Right
Fair comparison means acknowledging strengths. Cohere Command A+ has genuine advantages:
Native Citation Grounding
Every factual claim links to an explicit source span — not RAG-on-top, but trained into the model. This is a real differentiator for enterprise RAG applications and regulated document workflows.
48-Language Support
Up from 23 in Command A, this is the broadest multilingual coverage among the models compared here. If you need Romanian, Vietnamese, or Bengali, Cohere Command A+ has you covered.
Efficient Inference
375 tokens/second at W4A4, 113ms time-to-first-token. With only 25B active parameters, it's fast and lean. For high-throughput, low-latency tasks where reasoning depth matters less, this is an advantage.
But these strengths are narrow. Citation grounding matters for RAG, but the models Faraday shipped at publication outperformed it on the accuracy benchmark that underpins citation quality. 48-language support is impressive, but Qwen3.8-27B is built for multilingual work and runs on a single Faraday unit. Fast inference is valuable, but DeepSeek V4.1 Flash (16B active at decode) and MiMo-V2.6-Flash (15B active) activate fewer parameters per token than Command A+'s 25B.
The On-Premises Reality: Choice Matters
Both Cohere Command A+ and the Faraday Machines models can run on-premises. Cohere Command A+'s Apache 2.0 license is fully permissive — no revenue caps, no use-case restrictions. That's genuinely good for the ecosystem. But on-premises is only half the equation. The other half is which model you run:
| Consideration | Cohere Command A+ On-Premises | Faraday Machines On-Premises |
|---|---|---|
| Intelligence Index | Intelligence Index 37 at launch (pre-v4.3) — an estimated 13 on v4.3.2, per Artificial Analysis | Intelligence Index v4.3: GLM-5.3-Flash 42, MiMo-V2.6-Flash 38; Kimi K3 44 and GLM-5.3 45 on larger setups |
| Coding benchmarks | No SWE-bench or LiveCodeBench scores published at launch | SWE-bench Pro: Qwen3.8-27B 61.7%, MiniMax M3 59.0% (vendor-reported) |
| Agentic coding | 25% on Terminal-Bench Hard (agentic coding) | DeepSWE v1.1: DeepSeek V4.1 Flash 77.0 (DeepSeek-reported) |
| Hallucination rate | 86% non-hallucination — behind DeepSeek V4 Flash (96%) at launch | — |
| Context window | 128K input, 64K output context | Up to 1M context (DeepSeek V4.1 Flash, GLM-5.3-Flash, MiniMax M3) |
| Vendor lock-in | Locked to a single model from a single company — the last in the Command A family, with no successor yet | Choose the best model for each task — swap freely |
| Roadmap | Cohere's roadmap determines your capability | Your roadmap, your capability ceiling |
| Data residency | Canadian company, but US-hosted by default | Canadian data stays in Canada, on your hardware |
| Model lineup | — | Five open-weight models on standard tiers, from Starter up |
Running a model on-premises means controlling where your data goes. But it doesn't control what your AI can do. That's determined by which model you choose. Faraday Machines gives you five open-weight models on its standard tiers, more on larger setups — and the freedom to swap to whatever comes next.
What You're Really Giving Up
Choosing Cohere Command A+ because it's from a Canadian company is understandable. But consider what that choice costs you in capability:
A Reasoning Gap
At launch, the gap between Cohere Command A+ (37) and Kimi K2.6 (54) on the pre-v4.3 Intelligence Index was 17 points — the difference between a model that can handle basic tasks and one that can reason through complex, multi-step business problems. On today's v4.3.2 index, Artificial Analysis lists Command A+ at an estimated 13 (pending independent evaluation), against 42 for GLM-5.3-Flash. Treat the estimate with caution, but it's a capability ceiling, not a marginal gap.
No Coding Benchmarks
At launch, Cohere Command A+ had no published SWE-bench Verified or LiveCodeBench scores. If you're evaluating AI for software development, code review, or technical automation, this absence is a red flag. Faraday's current models publish coding results: DeepSeek V4.1 Flash on DeepSWE, Qwen3.8-27B and MiniMax M3 on SWE-bench Pro.
Vendor Lock-In
Cohere Command A+ is Cohere's only competitive open-weight model, and Cohere's changelog calls it the last model in the Command A family; as of October 2026, no successor has shipped. If you build your infrastructure around it, your capability ceiling is tied to Cohere's roadmap. Faraday Machines lets you swap between DeepSeek V4.1 Flash, GLM-5.3-Flash, Qwen3.8-27B, MiMo-V2.6-Flash and MiniMax M3 as your needs evolve — or as better models become available. Since this page was first published, Faraday's entire lineup has turned over.
Context Window Limitations
Cohere Command A+'s 128K input / 64K output context is adequate for most tasks, but DeepSeek V4.1 Flash and GLM-5.3-Flash offer 1M-token context windows, and MiniMax M3 up to 1M. For document analysis, codebase reasoning, and long-form generation, the difference between processing 128K and 1M tokens is transformational.
Frequently Asked Questions
Is Cohere Command A+ better than open models like GLM-5.3-Flash?
Not on Artificial Analysis’s current index. On Intelligence Index v4.3.2, Artificial Analysis lists Cohere Command A+ at an estimated 13, pending independent evaluation, while GLM-5.3-Flash scores 42 and MiMo-V2.6-Flash 38. Command A+ still has real strengths, including native citation grounding and 48-language support, but on general reasoning and coding benchmarks the open models Faraday Machines ships score higher.
What was Cohere Command A+’s Intelligence Index score at launch?
Cohere Command A+ scored 37 on the Artificial Analysis Intelligence Index at its May 2026 launch, on the pre-v4.3 version of the index. At the time, Kimi K2.6 scored 54, GLM-5.1 51, Qwen 3.6 Plus 50 and DeepSeek V4 Flash 47. Artificial Analysis rescaled the index with v4.3 in September 2026, so those May scores can’t be compared with current ones.
Can Cohere Command A+ run on-premises?
Yes. Cohere Command A+ is a 218B-parameter mixture-of-experts model with 25B active parameters, released under the permissive Apache 2.0 license, so it can be self-hosted; quantized, it needs about two H100 GPUs. Faraday Machines doesn’t ship Command A+. It installs DeepSeek V4.1 Flash, GLM-5.3-Flash, Qwen3.8-27B, MiMo-V2.6-Flash and MiniMax M3 on Mac Studio units at the customer’s site.
Is there a successor to Cohere Command A+?
No successor had shipped as of October 2026. Cohere’s changelog describes Command A+, released in May 2026, as the last model in the Command A family. Teams that build around a single vendor’s model are tied to that vendor’s roadmap. Faraday Machines installs several open-weight models on hardware the customer owns, and its lineup can be swapped as newer models are released.
References
[1] Artificial Analysis. (2026). "Intelligence Index Leaderboard." Available at: artificialanalysis.ai
[2] Artificial Analysis. (2026). "Cohere launches open weights model Command A+." Available at: artificialanalysis.ai
[3] VentureBeat. (2026). "Cohere cracks lossless quantization and native citations with Command A+." Available at: venturebeat.com
[4] Artificial Analysis. (2026). "DeepSeek is back among the leading open weights models with V4 Pro and V4 Flash." Available at: artificialanalysis.ai
[5] BenchLM.ai. (2026). "Command A+ Benchmarks." Available at: benchlm.ai
[6] Particula Tech. (2026). "DeepSeek V4 vs Kimi K2.6 vs GLM-5.1: Open-Weight Coding Tested." Available at: particula.tech
[7] Lushbinary. (2026). "Best Open-Source LLMs for AI Agents: DeepSeek V4 vs Kimi K2.6 vs Qwen 3.6 vs GLM 5.1." Available at: lushbinary.com
[8] Cohere. (2026). "Command A+ changelog." Available at: docs.cohere.com
[9] Artificial Analysis. (2026). "Cohere models." Available at: artificialanalysis.ai
[10] Artificial Analysis. (2026). "Artificial Analysis Intelligence Index v4.3." Available at: artificialanalysis.ai
[11] Artificial Analysis. (2026). "GLM-5.3-Flash." Available at: artificialanalysis.ai
[12] Artificial Analysis. (2026). "MiMo-V2.6-Flash." Available at: artificialanalysis.ai
[13] VentureBeat. (2026). "DeepSeek V4.1 Flash debuts with $0.003/1M off-peak cached input rate." Available at: venturebeat.com
[14] modelfit.io. (2026). "DeepSeek V4.1 Flash Mac memory requirements." Available at: modelfit.io
[15] Unsloth. (2026). "DeepSeek-V4.1-Flash-1017-GGUF." Available at: huggingface.co
[16] Unsloth. (2026). "GLM-5.3." Available at: unsloth.ai
[17] Unsloth. (2026). "GLM-5.3-Flash." Available at: unsloth.ai
[18] Unsloth. (2026). "Kimi-K3-GGUF." Available at: huggingface.co
Move Beyond Cohere Command A+
DeepSeek V4.1 Flash, GLM-5.3-Flash, Qwen3.8-27B, MiMo-V2.6-Flash and MiniMax M3 — with Kimi K3 and GLM-5.3 on larger setups — all on Faraday Machines hardware, on-premises, with no data leaks and no vendor lock-in.
Schedule Consultation