DeepSeek V3.2 Complete Guide: Speciale & Reasoning
DeepSeek V3.2 and V3.2-Speciale guide: 685B MoE with sparse attention, MIT license, $0.28/$0.42 API pricing and gold-level IMO and IOI 2025 results.
Key Takeaways
Corrected October 21, 2026: we re-checked this guide against DeepSeek’s release notes, pricing page, model cards and V3.2 technical report. The benchmark tables (which carried DeepSeek-V3 and R1 figures from December 2024 and January 2025), a non-existent cheaper V3.2 price tier, the R1 comparison, Speciale’s description, the IMO context and the hardware figures were corrected; unsourced cost-saving percentages, instruction-following scores and client examples were removed.
Update, October 2026: DeepSeek released V4 in preview on April 24, 2026 (V4-Pro, 1.6T total and 49B active parameters; V4-Flash, 284B and 13B; 1M-token context; MIT weights). From that date the deepseek-chat and deepseek-reasoner names that served V3.2 pointed to V4-Flash until DeepSeek retired both after July 24, 2026, as its change log had announced. In October 2026 the API serves deepseek-flash (V4.1-Flash, released September 10, 2026) and deepseek-v4-pro. Speciale’s temporary endpoint closed on December 15, 2025. The V3.2 and Speciale weights remain on Hugging Face.
DeepSeek V3.2 is an open-weight model released under the permissive MIT license on December 1, 2025. DeepSeek reports reasoning-benchmark performance comparable to GPT-5, at API prices of $0.28 per million input tokens (cache miss) and $0.42 per million output tokens. Built by DeepSeek, a Chinese AI research lab, this model challenges the assumption that state-of-the-art AI capabilities require proprietary, closed-source systems with expensive API access.
The release is complemented by V3.2-Speciale, a high-compute reasoning variant. DeepSeek reports that it scored 35 out of 42 on the 2025 International Mathematical Olympiad problems, which equals the IMO 2025 gold cutoff, and gold-level results at CMO 2025, IOI 2025 and the 2025 ICPC World Finals. This combination of a general-purpose model and a maximum-reasoning variant shows how close open-weight AI has come to commercial alternatives, while offering enterprises control, customization and cost advantages. This guide to the December 1, 2025 release explores DeepSeek V3.2’s capabilities, Speciale’s reasoning results, deployment considerations, and how organizations can use these models in production applications.
Open Source Advantage: Unlike proprietary models with usage restrictions and per-token pricing, DeepSeek V3.2’s MIT license allows unlimited commercial use, modification and self-hosting, which suits high-volume applications where API costs would be prohibitive.
What is DeepSeek V3.2?
DeepSeek V3.2 is a large language model developed by DeepSeek that uses a mixture-of-experts (MoE) architecture to achieve efficient, high-performance inference. Unlike dense models that activate all parameters for every token, MoE architectures route each token to a few specialized experts: V3.2 activates about 37B of its 671B main-model parameters per token, roughly 5.5%. According to DeepSeek’s V3.2 technical report, its only architectural change from DeepSeek-V3.1-Terminus is DeepSeek Sparse Attention (DSA): a lightweight “lightning indexer” picks 2,048 earlier tokens for each query token to attend to, cutting core attention cost from quadratic in sequence length to roughly linear and lowering long-context serving costs.
Training Efficiency: The widely quoted $5.5 million figure belongs to DeepSeek-V3, not V3.2. The DeepSeek-V3 technical report put V3’s final training run at 2.788M H800 GPU hours, or $5.576M at an assumed $2 per GPU hour, excluding prior research and ablations. V3.2 builds on that base with FP8 training and Multi-Token Prediction, then adds continued training and a reinforcement-learning budget that DeepSeek says exceeds 10% of pre-training cost.
Performance Benchmarks vs Competitors
All figures below are DeepSeek-reported, from Tables 2 and 3 of the V3.2 technical report (December 2025); DeepSeek ran its own models at temperature 1.0 with a 128K context. V3.2 is shown in thinking mode. DeepSeek’s tables give no SWE-bench Verified score for Speciale, which does not support tool calls.
| Benchmark | DeepSeek V3.2 (Thinking) | V3.2-Speciale | GPT-5 (High) | Claude Sonnet 4.5 |
|---|---|---|---|---|
| AIME 2025 | 93.1% | 96.0% | 94.6% | 87.0% |
| HMMT Feb 2025 | 92.5% | 99.2% | 88.3% | 79.2% |
| GPQA Diamond | 82.4% | 85.7% | 85.7% | 83.4% |
| LiveCodeBench (Coding) | 83.3% | 88.7% | 84.5% | 64.0% |
| Codeforces (Rating) | 2386 | 2701 | 2537 | 1480 |
| SWE-bench Verified | 73.1% | Not reported | 74.9% | 77.2% |
| Humanity’s Last Exam (text, no tools) | 25.1% | 30.6% | 26.3% | 13.7% |
Key Takeaway: In DeepSeek’s own tests, V3.2’s thinking mode sits close to GPT-5 (High) on math and coding, Speciale leads this group on most reasoning rows, and Claude Sonnet 4.5 leads on SWE-bench Verified. Gemini 3.0 Pro, also in DeepSeek’s tables, outscored V3.2 on every one of these rows and outscored Speciale on GPQA Diamond, LiveCodeBench, Codeforces and Humanity’s Last Exam. Choose your model based on your primary use case.
Speciale: Mathematical Reasoning Powerhouse
DeepSeek-V3.2-Speciale is a high-compute variant of V3.2 with the same architecture, built to test how far extended thinking can go. DeepSeek’s technical report says it was trained only on reasoning data with a reduced length penalty during reinforcement learning, and that DeepSeek added the dataset and reward method from DeepSeekMath-V2 to strengthen mathematical proofs. It is not a math-only model: it posts the highest Codeforces rating in DeepSeek’s tables after Gemini 3.0 Pro. The trade-offs are token-hungry answers and no tool calls. DeepSeek offered it only through its API (not its app or web chat), for community evaluation, on a temporary endpoint at V3.2’s prices, until December 15, 2025, 15:59 UTC, and published the weights on Hugging Face under MIT.
IMO Performance: 35/42 Points
DeepSeek reports that Speciale scored full marks on five of the six IMO 2025 problems and nothing on the sixth, for 35 out of 42. It also reports 102/126 at CMO 2025, 492/600 at IOI 2025 (10th place) and 10 of 12 problems at the 2025 ICPC World Finals (2nd place). The context for the IMO result:
- IMO Gold Medal Threshold: According to the IMO’s official results, the 2025 gold cutoff was 35 points, so Speciale’s score sits exactly on the line. Gold cutoffs from 2021 to 2025 ranged from 24 to 35 points.
- Human Comparison: At IMO 2025, 72 of 630 contestants (about 11%) won gold, and the mean score was about 18 of 42 by our calculation from the official individual results.
- Problem Complexity: IMO problems require deep mathematical insight, creative problem-solving approaches, and multi-step reasoning - not just calculation or formula application.
- AI Comparison: These were DeepSeek’s own runs after the contest, using a generate-verify-refine loop, with the final submissions published for others to check. By contrast, Google DeepMind’s Gemini Deep Think had its 35/42 graded by IMO coordinators in July 2025.
Real-World Applications
DeepSeek published no evaluations outside competitions and benchmarks, but Speciale’s results point to these candidate uses (no tool calls, so it suits standalone reasoning steps):
Quantitative Finance
Derivatives pricing, risk modeling, portfolio optimization, and algorithmic trading strategy development requiring complex mathematical formulations.
Scientific Research
Physics simulations, chemistry computations, engineering calculations, and mathematical proof assistance for research publications.
Educational Technology
Advanced tutoring systems that explain solution strategies, generate practice problems at appropriate difficulty levels, and provide step-by-step mathematical reasoning.
Data Science & ML
Statistical analysis, hypothesis testing, mathematical model development, and optimization problem formulation for machine learning pipelines.
DeepSeek V3.2 vs R1: Choosing the Right Mode
DeepSeek R1 (January 2025, updated as R1-0528 in May 2025) was built on the same 671B-parameter, 37B-active MoE base as DeepSeek-V3. Since V3.1 in August 2025, DeepSeek has folded reasoning into one hybrid model. Its API change log shows that from December 1, 2025 deepseek-chat ran V3.2’s non-thinking mode and deepseek-reasoner ran its thinking mode, so the API no longer served R1. The practical choice for V3.2 users is which mode to call; R1 remains an open-weight option for self-hosting.
| Feature | DeepSeek V3.2 | DeepSeek R1 (R1-0528) |
|---|---|---|
| Primary Focus | Hybrid: thinking and non-thinking modes, tool calls in both | Reasoning only, chain-of-thought |
| Response Style | Direct answers (non-thinking) or reasoning, then answer (thinking) | Shows thinking process, then answers |
| Speed | Fast in non-thinking mode | Slower (extended thinking phase) |
| Math (AIME 2025) | 93.1% (thinking) | 87.5% |
| Coding (SWE-bench Verified) | 73.1% | 57.6% |
| On DeepSeek’s API (Dec 2025) | Yes: deepseek-chat and deepseek-reasoner | No: open weights only |
| Best For | Production apps, agents, coding, content | Self-hosted reasoning on an earlier 2025 checkpoint |
Benchmark rows are DeepSeek-reported: V3.2 from its technical report, R1-0528 from its model card. The SWE-bench Verified scores used different agent frameworks (DeepSeek’s internal framework for V3.2, Agentless for R1-0528).
- Building production applications requiring speed
- Code generation and assistance tasks
- Content creation and summarization
- Customer support chatbots
- High-volume, cost-sensitive workloads
- Complex mathematical reasoning required
- Multi-step logical problem solving
- Agentic tasks that plan between tool calls
- Understanding the reasoning process matters
- Accuracy trumps response speed
Model Routing Strategy: Route simple queries to deepseek-chat and complex reasoning to deepseek-reasoner. Both bill at the same per-token rates, but thinking mode’s chain of thought counts toward output tokens (up to 64K per response versus 8K in non-thinking mode), so reasoning on every request raises costs and latency.
DeepSeek V3.2 API Pricing & Cost Optimization
DeepSeek offers some of the most competitive API pricing in the industry. Its pricing page as of December 4, 2025 listed one price for V3.2 in both modes and for Speciale, with no off-peak discount (DeepSeek ended those on September 5, 2025). The context window was 128K tokens; maximum output was 8K tokens in non-thinking mode, 64K in thinking mode and 128K for Speciale.
API Pricing Comparison
| Model | Input (Cache Hit) | Input (Cache Miss) | Output |
|---|---|---|---|
DeepSeek V3.2 (deepseek-chat / deepseek-reasoner) | $0.028/M | $0.28/M | $0.42/M |
| DeepSeek V3.2-Speciale (temporary endpoint to Dec 15, 2025) | $0.028/M | $0.28/M | $0.42/M |
| GPT-5.1 | $0.125/M | $1.25/M | $10/M |
| GPT-4 Turbo | N/A | $10/M | $30/M |
| Claude Sonnet 4.5 | $0.30/M | $3/M | $15/M |
Sources: DeepSeek’s pricing page (above); OpenAI’s GPT-5.1 and GPT-4 Turbo model pages; Anthropic’s pricing page (Claude Sonnet 4.5 cache reads). V3.2-Exp introduced these DeepSeek prices on September 29, 2025, cutting V3.1’s $0.07, $0.56 and $1.68 by 60%, 50% and 75%.
Cost Saving Opportunity: With cache hits at just $0.028/M tokens, 90% below the cache-miss rate, reusing the same prompt prefix cuts input costs sharply. DeepSeek’s caching is automatic and best-effort: it matches repeated prefixes in 64-token units, so structure prompts with the static parts first.
Self-Hosting vs API: Cost Comparison
No published benchmark gives a reliable self-hosting saving for V3.2, and the API’s list prices set a low bar. These are the inputs that decide it:
| Cost input | DeepSeek API | DeepSeek Self-Hosted |
|---|---|---|
| Price basis | $0.28/M input (miss), $0.42/M output | GPU hours, power, staff time |
| Weights to serve | None | About 690 GB (FP8, as released) |
| Minimum hardware | None | One eight-GPU node with H200-class memory (1,128 GB) |
| When it wins | Variable or modest volume | Sustained high utilization, or data that must stay in-house |
Cost Optimization Strategies
Structure prompts to maximize cache hits. Cache hits cost 10x less than cache misses, providing up to 90% savings on repeated system prompts.
Route simple queries to non-thinking mode and complex reasoning to thinking mode. The rates are the same, but thinking mode spends more output tokens, so reserve it for tasks that need it.
For non-real-time workloads, batch requests to maximize GPU utilization. The vLLM paper reports 2-4x higher throughput than earlier serving systems at the same latency.
V3.2 already ships in 8-bit FP8, half the memory of 16-bit weights. Going to 4-bit halves it again, at a quality cost DeepSeek has not measured, so benchmark your own tasks first.
Cost Factors to Consider
- GPU Compute: Hourly rental of an eight-GPU H200-class node from a cloud provider, or a one-time hardware purchase for on-premise deployment
- Engineering Overhead: Initial setup, optimization, monitoring, and maintenance require ML engineering resources
- Infrastructure Management: Load balancing, auto-scaling, monitoring, logging, and security hardening
- Model Updates: Periodically updating to newer DeepSeek versions or fine-tuning for specific domains
Deployment Guide: Running DeepSeek V3.2 in Production
Deploying DeepSeek V3.2 for production workloads requires careful infrastructure planning and optimization. Here are the main deployment options:
One-command deployment for local testing and development. Best for trying distilled models on consumer hardware.
- Simple setup
- Cross-platform
- Limited to smaller models
Production-ready with continuous batching, PagedAttention, and optimized throughput for cloud deployments.
- High throughput
- Production-ready
- Easy scaling
High availability with auto-scaling, load balancing, and enterprise-grade monitoring. AWS/Azure/GCP supported.
- Full model support
- Auto-scaling
- High availability
Distilled Models for Consumer Hardware
Can’t run the full model? DeepSeek released no smaller V3.2 variants, but its January 2025 R1 distills (Qwen and Llama models fine-tuned on R1 reasoning data) retain much of R1’s reasoning at a fraction of the size. Memory figures below are weight sizes at the stated precision, before KV cache:
| Model | Parameters | VRAM Required | Hardware Example |
|---|---|---|---|
| R1-Distill-Qwen-7B | 7B | ~8GB (8-bit) | RTX 3080, M1 Pro |
| R1-Distill-Qwen-14B | 14B | ~16GB (8-bit) | RTX 4090, M2 Max |
| R1-Distill-Llama-70B | 70B | ~40GB (4-bit) | A6000, 2x RTX 4090 |
| Full V3.2 (FP8, as released) | 685B | ~690GB | 8x H200 141GB (1,128GB) |
Step 1: Infrastructure Selection
Choose between cloud and on-premise deployment:
- Cloud (AWS, Azure, GCP): Lower upfront cost, easy scaling, pay-as-you-go. Best for variable workloads or initial testing. The FP8 weights need more than the 640GB of an eight-GPU A100 or H100 80GB node, so use eight-GPU H200 instances such as AWS p5e or p5en (1,128GB), Azure ND H200 v5, or Google Cloud A3 Ultra.
- On-Premise: Higher upfront investment in an eight-GPU server, but lower long-term costs for sustained high-volume usage. Full control over data residency and security. Best for consistent, high-volume workloads or strict compliance requirements.
Step 2: Model Optimization
Optimize model size and inference speed:
- Quantization: DeepSeek releases V3.2 in FP8, so the weights are already 8-bit. Community 4-bit builds halve memory again; DeepSeek publishes no quality figures for them, so evaluate on your own tasks before production use.
- Inference Framework: Use vLLM or SGLang, the two serving frameworks DeepSeek’s repository documents for the V3.2 architecture, for optimized serving with features like continuous batching and prefix caching.
- Caching: Implement KV cache optimization and prefix caching for repeated queries to cut prefill work and latency on shared prompts.
Step 3: Serving Infrastructure
Build production-ready serving infrastructure:
- Load Balancing: Distribute requests across multiple GPU instances using NGINX, HAProxy, or cloud load balancers for high availability.
- Auto-Scaling: Configure horizontal scaling based on request queue depth or GPU utilization metrics.
- API Gateway: Implement rate limiting, authentication, request logging, and monitoring using API gateways like Kong or AWS API Gateway.
- Monitoring: Deploy comprehensive monitoring with Prometheus/Grafana tracking latency, throughput, GPU utilization, error rates, and costs.
Step 4: Fine-Tuning (Optional)
Customize DeepSeek V3.2 for your specific use case:
- Domain Adaptation: Fine-tune on industry-specific data (legal documents, medical records, financial reports) to improve accuracy for specialized terminology and reasoning patterns.
- Instruction Tuning: Train the model to follow your preferred output format, tone, and style guidelines.
- Parameter-Efficient Fine-Tuning: Use LoRA or QLoRA to fine-tune only a small subset of parameters. The LoRA paper reports 10,000x fewer trainable parameters and one-third the GPU memory of full fine-tuning for GPT-3 175B.
When NOT to Use DeepSeek V3.2: Honest Guidance
While DeepSeek V3.2 is powerful and cost-effective, it's not the right choice for every use case. Here's honest guidance on when to consider alternatives.
- Strict regulatory compliance - HIPAA, certain government contracts, or industries requiring established audit trails
- Knowledge-heavy tasks - DeepSeek’s technical report says V3.2’s breadth of world knowledge still lags leading proprietary models
- No ML engineering resources - Self-hosting requires expertise your team may lack
- Data sovereignty concerns - When Chinese origin is a disqualifier for your organization
- Claude 4.5 - Strong instruction following and long-context analysis for complex agentic tasks; led SWE-bench Verified in DeepSeek’s own table
- GPT-5 - Reliable across diverse tasks, mature tool use, established enterprise support
- Azure OpenAI - Enterprise compliance, SOC 2, HIPAA-eligible with Microsoft backing
- DeepSeek V3.2-Speciale - Maximum reasoning for hard math and competitive programming, without tool calls
Data Privacy Consideration: While self-hosting ensures data never leaves your infrastructure, evaluate whether your organization’s policies allow use of Chinese-developed AI models. Self-hosting the open weights, rather than calling DeepSeek’s hosted API, is one way to address these concerns.
Common Mistakes with DeepSeek V3.2
Here are common mistakes to avoid when adopting DeepSeek V3.2 - and how to fix them.
The Error: Sending complex mathematical reasoning or multi-step logical problems to deepseek-chat when V3.2’s thinking mode would be more effective.
The Impact: Suboptimal results on reasoning tasks, user frustration, and wasted compute on retries.
The Fix: Implement mode routing - use deepseek-chat for speed-sensitive general tasks and deepseek-reasoner for complex reasoning. A simple classifier can route queries appropriately.
The Error: Not structuring prompts to take advantage of DeepSeek's prompt caching, paying full price for repeated system prompts.
The Impact: Up to 10x higher cost than necessary on the repeated part of each prompt, especially for applications with long, consistent system prompts.
The Fix: Structure prompts with static system instructions first, user content last. Cache hits cost $0.028/M vs $0.28/M for misses - a 90% savings.
The Error: Using aggressive 4-bit quantization for quality-sensitive production tasks to save memory costs.
The Impact: Quality loss that DeepSeek has not measured and that can compound in complex tasks, leading to user complaints and decreased trust.
The Fix: Start with the FP8 weights DeepSeek releases. Only drop to 4-bit if memory-constrained AND your own evaluations show your use case tolerates the difference.
The Error: Treating self-hosting as "just deploying a Docker container" without proper production infrastructure planning.
The Impact: Downtime during traffic spikes, poor latency, scaling issues, and ultimately higher costs than using the API.
The Fix: Plan for load balancing, auto-scaling, monitoring (Prometheus/Grafana), failover, and GPU memory management from day one. Budget for ML engineering time.
DeepSeek V3.2 vs GPT-5 vs Claude 4.5: Competitive Comparison
How does DeepSeek V3.2 stack up against the leading proprietary models of December 2025? Here’s a comparison to help you choose, using list prices and DeepSeek’s own benchmark table. If your workloads center on coding, it’s also worth weighing another open-weight option in our breakdown of GLM 4.6 against Claude Sonnet.
| Factor | DeepSeek V3.2 | GPT-5 / GPT-5.1 | Claude Sonnet / Opus 4.5 |
|---|---|---|---|
| Input Cost/M (cache miss) | $0.28 | $1.25 | $3-5 |
| Output Cost/M | $0.42 | $10 | $15-25 |
| Self-Hosting | Yes (MIT License) | No | No |
| Coding (SWE-bench Verified) | 73.1% | 74.9% (GPT-5 High) | 77.2% (Sonnet 4.5) |
| Tool Use (τ²-Bench) | 80.3% | 80.2% (GPT-5 High) | 84.7% (Sonnet 4.5) |
| Best For | High-volume, coding, cost-sensitive | General reliability, tool use | Agents, long-context, instruction |
Prices are list prices from DeepSeek’s, OpenAI’s and Anthropic’s pricing pages in December 2025. Benchmark rows are as DeepSeek reports them in Table 2 of its V3.2 technical report (V3.2 in thinking mode). Not every rival figure is a DeepSeek run: Sonnet 4.5’s 77.2% on SWE-bench Verified is Anthropic’s own published score. GPT-5 (High) trails on τ²-Bench but leads MCP-Mark in the same table (50.9%, against 38.0% for V3.2 and 33.3% for Sonnet 4.5).
- High-volume applications
- Cost is primary concern
- Data must stay on-premise
- Coding-focused workloads
- Need enterprise support
- Tool use is critical
- Compliance requirements
- Diverse task reliability
- Building AI agents
- Long-context analysis
- Precise instruction following
- Complex multi-step tasks
Use Cases and Applications
DeepSeek V3.2's combination of strong performance, MIT license, and cost efficiency makes it ideal for specific applications:
Companies processing millions of customer inquiries monthly pay $0.42 per million output tokens on DeepSeek’s API, against $10 for GPT-5.1 and $15 for Claude Sonnet 4.5 at December 2025 list prices.
Tip: Keep the system prompt and knowledge base text identical across requests so they bill at the cache-hit rate.
Document analysis, code assistance, knowledge base Q&A for employees. Self-hosting ensures proprietary data and trade secrets remain internal.
Benefit: Self-hosted weights keep code and documents inside your own infrastructure.
Universities can leverage Speciale’s open weights for mathematical research assistance without per-token costs constraining experimentation.
Benefit: DeepSeek published Speciale’s IMO, CMO, IOI and ICPC submissions, so researchers can check its reported results themselves.
SaaS companies can embed AI capabilities directly into products without per-user API costs eating into margins.
Benefit: MIT license allows product integration without royalty restrictions.
Conclusion
DeepSeek V3.2 and Speciale represent a breakthrough in open-weight AI: reasoning scores that DeepSeek puts close to GPT-5, complete deployment flexibility, and API prices well below commercial alternatives ($0.42 per million output tokens against $10 for GPT-5.1). The MIT license removes barriers that have historically limited open-source AI adoption, enabling enterprises to self-host, modify, and deploy without restrictions or per-token fees.
For organizations with high-volume AI workloads, strict data residency requirements, or specialized domain needs, DeepSeek V3.2 offers a compelling alternative to proprietary APIs. Combined with Speciale’s gold-level competition results, these models show open-weight AI matching commercial offerings on many reasoning benchmarks in DeepSeek’s own tests. DeepSeek itself says V3.2 still trails models such as Gemini 3.0 Pro on world knowledge, token efficiency and the hardest tasks, while offering control, customization and cost advantages that proprietary systems cannot deliver.
Bottom Line: DeepSeek V3.2 is a value leader - delivering near-frontier results in coding and math at a fraction of competitor costs. It’s not trying to replace GPT-5 or Claude as a universal brain, but excels where most developers actually work.
Deploy Open Source AI at Scale
We help organizations deploy and optimize open-source AI models like DeepSeek V3.2, from infrastructure planning to production deployment and fine-tuning.
Frequently Asked Questions
Deep dives on AI, marketing and development.
Practical guides and fresh insights by email. No recycled takes.
Related Guides
Continue exploring open-source AI models and deployment strategies.