Command R 35B

Positioning

Cohere Command R 35B is the smaller sibling of Command R+ 104B — Cohere's open-weight retrieval-augmented generation model at the 35B parameter dense tier. Released under CC-BY-NC-4.0 (research/non-commercial). Like Command R+, it's explicitly trained for RAG with first-class document citation, multilingual coverage (10+ languages), and tool-use support. The 35B parameter count makes it dramatically cheaper to serve than the 104B variant — fits a single 48 GB workstation card or ~24 GB at Q4-Q5 quantization.

Strengths

• Retrieval grounding with citation accuracy. Same architecture-level training as Command R+ — citation accuracy is meaningfully better than equal-parameter Llama 3 / Qwen 3 models.
• Multilingual at the 35B tier. Strong on Arabic, Korean, Hebrew, Vietnamese, and other Cohere-prioritized languages.
• Tool-use support is real. Function calling + multi-step tool use trained as first-class capability.
• 128K context with stable degradation curve.
• Practical hardware fit. 35B FP16 = ~70 GB; 35B Q4-Q5 = ~22-26 GB on single 24 GB GPU.

Limitations

• License is non-commercial. CC-BY-NC-4.0 — production commercial deployment requires Cohere licensing. Same constraint as Command R+.
• Reasoning trails Llama 3.1 70B / Qwen 3 32B. The retrieval-and-multilingual focus trades general reasoning for specialized capability.
• English-only quality is below similar-size general-purpose models.
• Smaller than Command R+ on multilingual depth. Best multilingual + RAG goes to Command R+ 104B; R 35B is the cheaper-to-serve alternative.
• Updates are slower than Llama / Qwen / DeepSeek. Cohere's release cadence is much less frequent.

Real-world performance

• vs Command R+ 104B: Command R+ wins clearly on most metrics, but R 35B serves at ~1/3 the inference cost — meaningful at scale.
• vs Llama 3.1 70B: Llama wins for English-only general tasks. Command R 35B wins on multilingual + retrieval citation.
• vs Qwen 3 32B: Qwen 3 is stronger on most general benchmarks. Command R 35B wins on retrieval citation accuracy.
• vs Llama 3.1 8B: Larger and more capable across the board, but ~5× the inference cost.

Should you run this locally?

Yes if you specifically need 35B-class multilingual + RAG citation for research / non-commercial deployments, you have 24-48 GB GPU capacity, and Command R+ 104B is overkill / too expensive to serve.

No if you need permissive commercial licensing (pick Llama 3.1 70B or Qwen 3 32B), reasoning-heavy workloads (pick DeepSeek/Qwen 3), or English-only workflows.

How it compares

• vs Command R+ 104B: R+ is the larger sibling with stronger multilingual + RAG capabilities at 3× the inference cost.
• vs Aya Expanse 32B: Both are Cohere multilingual models with non-commercial licenses. Aya emphasizes language coverage; Command R 35B emphasizes RAG citation. Different specializations.
• vs Llama 3.1 70B: Llama is permissive license + stronger general; Command R 35B is RAG-citation-tuned.

Run this yourself

• Single 24 GB GPU at Q4-Q5: RTX 4090, RTX 5090, used 3090.
• Single 48 GB GPU at FP16: RTX 6000 Ada, L40S.
• Apple Silicon: Mac Studio M3 Ultra or MacBook Pro M4 Max at FP16.
• vLLM serving: vllm serve CohereForAI/c4ai-command-r-v01 --max-model-len 32768.
• Cloud rental: Runpod / Lambda RTX 4090 ~$0.50-0.70/hr.