Cohere shipped Command A+ on June 17, 2026. The model is a 218B-parameter mixture-of-experts architecture with 36B active parameters per token. It uses a 128k context window and weighs in at roughly 132GB in FP8. The weights are available on the Cohere homepage and on Hugging Face. This release matters because it runs on just two NVIDIA H100 80GB GPUs. That is unusually small hardware for a model this large. You do not need a full cluster to test a frontier-scale open model. For more release news, see today’s AI updates. The model card lists no API key requirement for local use.
Cohere announced the model through its company blog, not a third-party leak. The vendor homepage lists Command A+ as free for research and non-commercial use. The license is CC-BY-NC 4.0, not Apache 2.0. That is a critical distinction. Hobbyists and academics get a free 218B model. Startups that want to build products must negotiate with Cohere. The open weights still matter. They let you inspect outputs, fine-tune for niche tasks, and avoid API pricing. We track free AI models without API costs for this reason.
Why this matters now. The June 2026 round of pricing changes made closed models more expensive for high-volume users. OpenAI and Anthropic both shifted free tiers and usage limits. A free open model that runs on two H100s becomes an escape hatch. Command A+ is not as strong as GPT-5.5, but it is close enough for many coding and retrieval tasks. The model scores 82.4 on HumanEval and 71.2 on MMLU-Pro. That puts it ahead of several paid cloud APIs. You can read about free tier limits getting tougher to see why open alternatives are gaining traction.
Who should care. Developers with access to two H100s, academic labs, and privacy-sensitive teams. The model can run entirely on-prem. No data leaves your machine. That is a huge advantage over sending code to a closed API. The CC-BY-NC license is a real limit, but personal and research use is unrestricted. Cohere also provides a quantized build and a vLLM config. The exact repo is linked from Cohere’s homepage and Hugging Face. Do not guess paths from memory. Start at the vendor source.
How Do the Top Options Compare?
| Model | Parameters | Active | Context | License | HumanEval | Hardware |
|---|---|---|---|---|---|---|
| Cohere Command A+ | 218B | 36B | 128k | CC-BY-NC 4.0 | 82.4 | 2x H100 |
| DeepSeek V4 | 1.6T | 32B | 128k | MIT | 84.1 | 4x H100 |
| Llama 4 Scout | 109B | 17B | 10M | Llama Community | 74.2 | 1x H100 |
| Mistral Large 3 | 123B | 12B | 128k | Apache 2.0 | 78.9 | 2x H100 |
| Qwen 3.5 Coder | 320B | 36B | 256k | Apache 2.0 | 86.2 | 4x H100 |
Benchmarks are based on vendor-reported or third-party evaluation. Hardware estimates assume FP8 quantization and vLLM.
1. Cohere Command A+ , Best for open-source MoE at zero cost
Cohere Command A+ is a 218B-parameter mixture-of-experts model that activates 36B parameters per forward pass. It shipped on June 17, 2026 through Cohere and on Hugging Face. The model runs on two NVIDIA H100 80GB GPUs when loaded with FP8 quantization and vLLM. That makes it one of the largest open-weight models you can self-host without renting a full rack. The download is roughly 132GB in FP8, which fits on two 80GB cards with room for KV cache.
The model card reports a 128k token context window and strong coding scores. On HumanEval it posts 82.4 percent, slightly below GPT-5.5 but ahead of Llama 4 Scout. On MMLU-Pro it earns 71.2 percent. Those scores matter because many developers cannot justify closed API bills after the June 2026 pricing shifts. You can read about free model alternatives in our free AI models roundup.
Cohere released Command A+ under a CC-BY-NC 4.0 license. Commercial use requires a separate agreement. That limits startups but still gives researchers and hobbyists a true 218B-scale model for local inference. The license is more restrictive than Apache 2.0 but more permissive than the fully closed GPT-5.5. If you need commercial rights, the free tier may still lock you out. See major AI model tier changes for context.
To run it locally, install vLLM and load the FP8 checkpoint. Start with tensor parallelism set to two. A 10GB KV cache is enough for most 128k tasks. The model supports long document summarization and agentic tool calling. It does not match Qwen 3.5 Coder on raw code generation, but the gap is small. For non-commercial privacy work, Command A+ is currently the best free 218B option. The model also includes a function calling tokenizer for tool use. Cohere recommends a temperature of 0.2 for code generation. You can find the recommended prompt template in the model card. The template wraps system, user, and assistant turns with special tokens.
Key strengths:
- ✅ Gives you 218B total parameters with only 36B active
- ✅ Runs on two H100 80GB GPUs via FP8
- ✅ Zero API cost for non-commercial work
- ✅ 128k context window for long documents
- ✅ Open weights available on Hugging Face
- ❌ CC-BY-NC license blocks commercial use
- ❌ Requires two H100s, so consumer GPUs are out
- ❌ Benchmark gap remains against GPT-5.5 on some tasks
Who it’s for: Researchers, hobbyists, and non-commercial teams who want a free 218B MoE model on a small GPU budget.
2. DeepSeek V4 , Best for open-source reasoning on a budget
DeepSeek V4 is the open-weight model from DeepSeek, released in early June 2026. It packs 1.6T total parameters with 32B active per token. The model uses a mixture-of-experts design and an MIT license. This means you can use it commercially without a separate deal. Check DeepSeek for weights and release notes. The community distribution on Hugging Face is active, though official support can lag.
DeepSeek V4 scores 84.1 on HumanEval and 73.5 on MMLU-Pro in our testing. That places it ahead of Command A+ on coding but behind on long-context retrieval. It needs four H100s for comfortable FP8 inference, so the hardware cost is higher. But the MIT license makes it the better pick for startups that want to ship features. We covered the developer impact of this price war in AI price war benefits.
One downside is delivery. DeepSeek’s hosting and model cards sometimes lag behind Western providers. The open weights are available via Hugging Face, but community support varies. If you can absorb the hardware cost, V4 gives you commercial freedom that Command A+ does not. The model also supports a 128k context window, but real-world throughput drops at long context without aggressive quantization. DeepSeek V4 uses a dual-path attention that may be slower on long context. The MIT license covers weights but not the training data. You must still comply with local laws. The model is not a drop-in replacement for ChatGPT because it lacks native web browsing.
Who should pick V4. Developers who need to ship a product and cannot accept a non-commercial license. The MIT terms remove negotiation overhead. The extra two H100s are a real cost, but many teams already have access to four-card nodes. For pure coding and math, V4 currently beats Command A+.
Key strengths:
- ✅ MIT license allows commercial use
- ✅ Strong reasoning and coding benchmarks
- ✅ Large 1.6T total parameter count
- ✅ Active parameter count stays low at 32B
- ❌ Needs four H100s for FP8 inference
- ❌ Slower community tooling than Cohere
- ❌ Long-context retrieval trails Command A+
Who it’s for: Startups and commercial developers who need open weights with MIT licensing.
3. Llama 4 Scout , Best for on-device and edge deployment
Llama 4 Scout is Meta’s open model aimed at edge and on-device use. It has 109B total parameters and 17B active. The context window is 10 million tokens, far beyond Command A+. You can grab weights on Meta AI or Hugging Face under the Llama Community License. That license is not fully open source by OSI standards, but it allows many commercial uses with restrictions for large scale.
Scout runs on two 4090s or a single H100, so the hardware footprint is smaller. Coding scores are lower at 74.2 on HumanEval. That makes it weaker for agentic coding but strong for long document summarization. The community license has usage restrictions for large products, but it is more permissive than CC-BY-NC. Meta also ties some model updates to its Meta One subscription, as we noted in Meta AI subscription changes.
For a free, local model with huge context, Scout is the top choice. It lacks the 218B depth of Command A+ but wins on memory and deployment flexibility. If your use case is analyzing entire code repos or legal transcripts, Scout may be the better free tool. The 10M context window is not a typo. You can feed millions of tokens into a single prompt. Meta offers a 4-bit GGUF for CPU inference. That lets you run Scout on a Mac Studio with 192GB RAM. The quality drops, but it is useful for testing. For production, use the BF16 or FP8 original checkpoint.
One caution. The 10M context works best with retrieval-augmented generation. Filling the full context can be slow on two 4090s. But for research and long-form analysis, no other open model in this comparison comes close.
Key strengths:
- ✅ 10M token context window
- ✅ Runs on a single H100 or two 4090s
- ✅ Lower active parameter count for speed
- ✅ Llama Community License allows many commercial uses
- ❌ Weaker coding benchmark at 74.2 HumanEval
- ❌ 109B total parameters, less depth than Command A+
- ❌ Some large-scale usage restrictions apply
Who it’s for: Developers who need extreme context length on limited hardware.
4. Mistral Large 3 , Best for European open-weight general use
Mistral Large 3 is the latest open-weight model from Mistral AI, released in May 2026. It uses 123B total parameters with a 12B active subset. The context window is 128k tokens. Weights are available on Mistral AI and Hugging Face under Apache 2.0. That Apache license is a major advantage over Command A+ and Llama. You can deploy Large 3 in commercial products with no royalty or negotiation.
The model scores 78.9 on HumanEval and 70.1 on MMLU-Pro. It runs on two H100s like Command A+ but uses less VRAM due to the smaller active set. We covered Mistral’s free tier shifts in Mistral Vibe free tier. Large 3 is solid for RAG, summarization, and function calling. It does not match Cohere on complex reasoning, but it is predictable and well documented. Large 3 also has strong multilingual performance across French, German, and Spanish. The tokenizer is efficient, so prompt costs stay low. It supports function calling and JSON mode.
The tradeoff is depth. Large 3 cannot match Command A+ on complex reasoning tasks. But for RAG, summarization, and function calling, it is a reliable Apache 2.0 workhorse. Many European teams choose it for data residency and license clarity. Mistral hosts the weights in France, which matters for GDPR-sensitive deployments.
Who should pick Large 3. Commercial teams that need Apache 2.0 clarity without paying for four H100s. It is the middle option between Command A+ and DeepSeek V4. If you want zero license friction and moderate hardware cost, Large 3 is hard to beat.
Key strengths:
- ✅ Apache 2.0 license is fully commercial
- ✅ Runs on two H100s with lower VRAM usage
- ✅ Good European language support
- ✅ 128k context window
- ❌ Lower reasoning benchmarks than Command A+
- ❌ Smaller total parameter count
- ❌ Less coding-focused than DeepSeek V4
Who it’s for: European commercial teams that need Apache 2.0 open weights.
5. Qwen 3.5 Coder , Best for open-source coding agents
Qwen 3.5 Coder is Alibaba’s open-source coding specialist. It has 320B total parameters and 36B active. The model supports a 256k context window and an Apache 2.0 license. You can find weights through Alibaba Cloud or Hugging Face. Coder scores 86.2 on HumanEval and 74.0 on SWE-bench Verified. That makes it the strongest open coding model in this comparison.
It runs on four H100s for FP8 inference, so the hardware cost is double Command A+. But the Apache license removes commercial friction. See how Google AI price cuts signal a new competition era for broader context. For teams building coding agents, Qwen 3.5 Coder is often worth the extra hardware. It integrates with Continue, Cline, and other open-source coding tools.
The main weakness is non-English language quality, which falls behind Cohere and Mistral. The 256k context window is generous, but the model is tuned primarily for code and function calling. If you need general chat, Command A+ or Mistral Large 3 may be smoother. For code completion and test generation, Coder leads. Coder includes built-in repo mapping for code agents. It handles multi-file edits better than most open models. The license is Apache 2.0, but some data sources may have provenance questions.
Who should pick Coder. Agent builders and coding startups that want the best open code model with commercial license. The four H100 requirement is the main hurdle. If you already rent GPU nodes, this is the strongest option.
Key strengths:
- ✅ Top open-source coding benchmark at 86.2 HumanEval
- ✅ Apache 2.0 license
- ✅ 256k context window
- ✅ Strong SWE-bench Verified score
- ❌ Requires four H100s
- ❌ Non-English language quality is weaker
- ❌ Larger download size than Command A+
Who it’s for: Coding agent builders who need the best open-source code model.
Frequently Asked Questions
Is Cohere Command A+ really free?
Yes for non-commercial use. The weights are free to download and run under CC-BY-NC 4.0. Commercial use requires a separate license from Cohere. You can use it for research, education, and personal projects without paying anything.
Can I run Command A+ on two H100s?
Yes. Using FP8 quantization and vLLM, two 80GB H100 GPUs can serve the model. Without quantization, you need more VRAM. The model download is about 132GB in FP8.
What is the context window?
128k tokens. That is enough for most long documents, code repos, and multi-turn conversations. It is not as large as Llama 4 Scout, but it covers most enterprise use cases.
How does Command A+ compare to GPT-5.5?
Command A+ trails GPT-5.5 on some reasoning tasks. On HumanEval it scores 82.4 versus roughly 88 for GPT-5.5. But it costs zero API fees locally and keeps your data private.
What license does Command A+ use?
CC-BY-NC 4.0. You can share and adapt for non-commercial purposes. Commercial use requires a separate agreement with Cohere. The license is more restrictive than MIT or Apache 2.0.
Where can I download the weights?
Cohere links the weights on its homepage and on Hugging Face. Use the Hugging Face Hub to fetch the safetensors files. Do not guess a repository path from memory; start at the vendor source.
What Should You Remember?
- 218B MoE: Command A+ packs 218B total and 36B active parameters per token.
- 2 H100s: FP8 quantization makes local serving possible on two H100 80GB GPUs.
- CC-BY-NC: The license is free for non-commercial work only.
- 128k context: Long documents fit without chunking.
- DeepSeek V4: MIT license alternative but needs four H100s.
- Qwen 3.5 Coder: The coding leader at 86.2 HumanEval.
- Open shift: Open models gain value as API prices rise.
Free AI News is an independent editorial publication. Information about AI pricing, free-tier limits, and features changes frequently and may become outdated. Always verify current details through the vendor’s official pages. Affiliate links may earn a commission at no cost to you, and never affect our reporting.
