DeepSeek V3.2 is the December 2025 update to DeepSeek's frontier-class open-weight model — 685B total parameters with 40B active per token (Mixture of Experts), MIT-licensed, stronger coding and reasoning performance than V3, 914K downloads, 1,336 likes, and serves as the base for DeepSeek R1 reasoning distillations.
Why It's in Trial
DeepSeek V3.2 earns Trial as an incremental but meaningful upgrade to the proven V3 foundation:
- Proven architecture: V3/V3.2 establish DeepSeek as a top-tier open-weight provider (SWE-bench 71.6%, top coding performance among open models)
- MIT License — unrestricted commercial use; self-hostable weights on Hugging Face
- Latest base model: V3.2 (Dec 2025) succeeds V3 (Dec 2024); modest improvements in coding and reasoning beyond V3
- Efficient MoE design: 40B active parameters out of 685B total dramatically reduces inference cost vs. dense models
- Strong adoption: 914K downloads, 1,336 likes; ranked among top open-weight models
- Inference provider diversity: Novita and Fireworks AI support; multiple open-source inference frameworks (vLLM, SGLang)
- R1 foundation: DeepSeek R1-Distill models (1.5B, 7B, 32B, 70B) base on Qwen 2.5 or Llama, but V3.2 remains the undistilled frontier base
Positioned in Trial (not Adopt) because: while coding benchmarks place V3.2 in the frontier tier, it remains behind Claude Opus 4.6 and GPT-5.4 on general reasoning and GPQA tasks; broader adoption outside coding is still developing.
V3 vs. V3.2 vs. V3.1 Comparison
| Model | Parameters | Release | Status | SWE-bench | Context |
|---|---|---|---|---|---|
| V3 | 685B MoE (40B active) | Dec 2024 | Superseded | 71.6% | Original frontier release |
| V3.1 Terminus | 685B MoE (40B active) | Mar 2025 | Current variant | ~71% | Hybrid reasoning mode |
| V3.2 | 685B MoE (40B active) | Dec 2025 | Latest base | ~72% (est.) | Improved over V3 |
| V3-0324 | 685B MoE (40B active) | Mar 2025 | Minor update | Similar to V3 | Snapshot version |
Recommendation: Use V3.2 for latest performance; V3.1 Terminus for hybrid reasoning workloads; older versions (V3, V3-0324) are superseded.
Coding Performance
DeepSeek V3.2 leads open-weight models on multiple coding benchmarks:
| Benchmark | V3.2 (est.) | GLM-5 | Qwen-2.5-Coder-32B | Claude Opus 4.6 |
|---|---|---|---|---|
| SWE-bench Verified | ~72% | 77.8% | 36.5% | 74% |
| LiveCodeBench | ~50% (est.) | Data unavailable | 37.2% | — |
| HumanEval | ~90% | ~92% | ~85% | 92%+ |
Note: V3.2 is estimated slightly above V3 (71.6%) based on release notes; full independent benchmarking pending. GLM-5 remains highest-scoring open-weight on SWE-bench.
Deployment & Cost
Self-hosted:
- vLLM, SGLang, LM-Studio support
- ~330GB memory (BF16); ~165GB with FP8 quantization
- MoE routing overhead minimal (~5%) compared to dense models
- Latency comparable to 120B dense models due to 40B active parameters
Managed inference:
- Novita (live)
- Fireworks AI (live)
Cost context:
- Self-hosted on 2x H100 (80GB): ~$30K/month cloud cost + engineering overhead
- Managed API: ~$0.27/M input, $1.08/M output via Novita (variable)
Architecture & Training
- Mixture of Experts (MoE): 685B total, 40B active per token (similar to GLM-5's design)
- Context window: 128K tokens (standard for frontier open-weight)
- Training data: Unclear composition; DeepSeek maintains confidentiality
- Post-training: SFT + reinforcement learning for reasoning and alignment
- Quantization: FP8 support; dynamic quantization for mixed-precision inference
When to Choose V3.2 (vs. V3.1 Terminus)
- Coding focus — V3.2 has marginally better code performance than V3.1
- General-purpose inference — V3.2 is the straightforward choice for non-reasoning tasks
- Cost-critical deployments — MoE efficiency (40B active) reduces compute vs. dense models
- Self-hosting — weights fully open, MIT licensed, no API cost overhead
When to choose V3.1 Terminus instead:
- Explicit reasoning required — R1-style chain-of-thought via hybrid reasoning mode
- Complex problem-solving — multi-step reasoning tasks
Key Characteristics
| Property | Value |
|---|---|
| Total parameters | 685B (Mixture of Experts) |
| Active parameters | 40B per token |
| Context window | 128K tokens |
| Latest version | V3.2 (December 2025) |
| License | MIT |
| Provider | DeepSeek |
| Weights | Hugging Face: deepseek-ai/DeepSeek-V3.2 |
| Released | December 1, 2025 |