LangGraph
Production orchestration. Best fit for DeepSeek's context-management strategies.
DeepSeek V3.2 is the open-weight model that beat GPT-5 on the hardest reasoning benchmarks. Released December 1, 2025 under MIT license, it pairs a 671 billion total / 37 billion active MoE architecture with DeepSeek Sparse Attention, a fine-grained sparse pattern layered on Multi-head Latent Attention that cuts long-context inference cost while leaving output quality nearly unchanged.
Specs that matter when picking the model and the hardware.
671B total, 37B active per token
MoE, 256 routed + 1 shared expert, DeepSeek Sparse Attention
128K tokens (max output 64K)
MIT
December 1, 2025 (V3.2-Exp in September 2025)
V3.2 (general), V3.2-Speciale (deep reasoning, no tool calling)
376GB total, multi-GPU mandatory
~675GB, 8x H200 node
8-15 (single user)
200-400 batched
Three tiers, real SKUs, real throughput numbers.
Throughput: 8-15 tokens/sec. Best for: Single power user, research, occasional hard reasoning, not production serving.
Throughput: 30-60 tokens/sec. Best for: 5-15 user firms running hard problems, lab settings, code-heavy teams.
Throughput: 200-500 tokens/sec. Best for: Production serving, agentic workflows, 50+ concurrent reasoning users.
GGUF, AWQ, FP8, INT4.
FP8 native on Hopper and Blackwell for production. Q4_K_M for Mac single-user. Q2_K only if you must fit 192GB and accept a 5 to 8 percent quality dip on reasoning.
Reasoning models suffer more from aggressive quantization than chat models because chain-of-thought amplifies precision errors. Q4 holds up well on knowledge and edit-style code. Q2_K shows real wobble on math and structured reasoning. 1.58-bit demos exist (2x RTX 4090 plus RAM) but are not production-grade.
File size: 210GB.
File size: 376GB.
File size: ~675GB.
File size: 1.34TB.
DeepSeek Sparse Attention (DSA) gives near-linear scaling for long context, which is decisive on agentic loops where context grows with every tool call.
DeepSeek's official integration target. SWE-bench Verified 72 to 74 percent on Claude Code framework. Override OPENAI_BASE_URL to your vLLM. DeepSeek V3.2 is the open model that comes closest to Claude Sonnet on Claude Code's harness. Pineido's pick when the buyer wants Claude-Code-quality output without the API spend.
Officially benchmarked. The DeepSeek paper notes Roo Code simulates tool interactions via user messages, and DeepSeek V3.2's context-management strategy handles this correctly. SWE-bench Verified 72 to 74 percent on Roo Code framework. Multi-mode (Code, Architect, Ask) all functional.
OpenAI Compatible provider, vLLM endpoint. Watch the active vLLM bug where DeepSeek tool-call tags can leak into reasoning content on multi-turn loops. Recent vLLM releases include the fix (commit 36056). Pin vLLM >= 0.11.2.
Strong fit. Aider's edit format aligns with DeepSeek's instruction-following. Use --model openai/deepseek-v3.2 against your vLLM endpoint. The 128K context limit is the constraint, not the editing accuracy.
Container-based agent. DeepSeek V3.2 with Terminus in non-thinking mode hit 39.3 on Terminal Bench 2.0 per the paper. SWE-bench Verified 72 to 74 percent translates to strong autonomous-loop performance. Pineido's pick for repo-scale tasks when budget rules out frontier APIs.
config.yaml with provider: openai, apiBase to vLLM /v1, capabilities: [tool_use]. Agent mode auto-detects native tool calling. DeepSeek V3.2's reasoning parser separates thinking from response, so Continue's chat panel renders cleanly.
Officially supported per DeepSeek's own coding-agent integration guide. Drop-in OpenAI endpoint config. The Hermes Agent and Codex frameworks are also listed in DeepSeek docs.
Where this model + agent stack ships real work.
Feedback from the people who were in the room when decisions were made.
I used to spend every Sunday reconciling contractor invoices. Now they match automatically. I got my weekends back.
We stopped debating architecture in circles and shipped a working release that leadership could evaluate live.
They did not ask us to trust a rewrite plan. They proved one critical workflow first, then scaled from evidence.
Long-tail agent task generalization is DeepSeek V3.2's standout: 1,800 environments and 85K complex prompts in post-training.
Set OPENAI_BASE_URL to vLLM. DeepSeek V3.2's reasoning trace makes GPT Researcher's plan step noticeably better. Reports of 5K to 8K words with 15+ sources complete in 5 to 10 minutes.
Pineido default. DeepSeek V3.2's context-management strategies mesh with LangGraph's checkpoint-and-resume model. When tool history exceeds 80 percent of the 128K budget, LangGraph can swap in the Summary or Discard-75 strategy and continue cleanly.
Generic OpenAI provider. The reasoning parser separates thinking from final answer in AnythingLLM's chat UI. Great for buyer demos because the thinking trace is visible and explains the answer.
Tool calling and document upload work cleanly. DeepSeek's reasoning-mode toggle exposes via Open WebUI's per-conversation settings. Watch the 128K context limit on heavy document chats.
Sequential research crews. DeepSeek V3.2's instruction following is strong enough that CrewAI's role and goal text drives the agents reliably. Use thinking-mode for the researcher role, non-thinking for the writer role.
GroupChat works but watch token cost: every agent turn replays full history through DeepSeek's 128K window. Use Discard-75 context-management strategy to keep round counts bounded.
Best retrieval pairing. Use LlamaIndex query engines for indexing, then expose them as tools inside a LangGraph or DeepSeek-native agent. The reasoning parser surfaces retrieval-decision logic in the trace.
Where this model + research agent stack delivers reports your team actually uses.
Pick the framework by workflow shape.
Production orchestration. Best fit for DeepSeek's context-management strategies.
Sequential crews. Fast to demo with DeepSeek's strong instruction following.
GroupChat. Watch context budget on multi-agent debates with DeepSeek's 128K cap.
Prompt compilation. Pairs well with DeepSeek's reasoning trace for chain-of-thought optimization.
Type-safe agents. DeepSeek's structured output reliability fits clean schemas.
Retrieval layer. Hybrid stack with LangGraph orchestration on top.
Three buyer profiles where this model is the right answer.
Task: Strategy code, derivative pricing, complex math validation. Outcome: IMO/IOI gold-medal class reasoning runs in-house. MIT license clears the legal review with no caveats. 8x H100 ROI inside 6 months versus paid frontier API.
Task: Prior art analysis, claim drafting, technical specification review across 10+ technical domains. Outcome: Frontier reasoning quality on hard technical claims, full 128K context for multi-patent comparison. MIT license means no attribution friction in client deliverables.
Task: Molecular reasoning, protocol generation, deep literature synthesis. Outcome: Speciale variant delivers theorem-proving and deep-reasoning on par with closed labs. Hardware investment justified by frontier quality without API data exposure.
Where this model wins, where it loses.
DeepSeek V3.2 wins on hardest math, code, and IMO-class reasoning. Qwen3 235B wins on hardware floor (143GB versus 376GB at Q4) and on hybrid thinking-mode flexibility. Below 4x H100 budget, Qwen3 235B is the rational pick.
DeepSeek V3.2 performs comparably to GPT-5 on commodity tasks. V3.2-Speciale exceeds GPT-5 on theorem proving and IMO, scoring 96.0 on AIME 2025 versus GPT-5-High at 94.6. The catch is verbosity: Speciale spends far more output tokens to reach those scores. The win for V3.2: data stays on your hardware. The cost: enterprise-grade GPU capex.
Both are 670B-class MoE on 128GB+ hardware. Mistral Large 3 wins on multimodal vision. DeepSeek V3.2 wins on reasoning and code. Mistral Large 3 has a slightly better small-firm story via 14B/8B/3B siblings.
Day 4 vLLM serves V3.2 with the v32 tool-call parser plus reasoning parser.
vLLM serves DeepSeek V3.2 with --tensor-parallel-size 8 (or TP=2 with DP/EP for Hopper to avoid head-padding penalty), --tokenizer-mode deepseek_v32, --tool-call-parser deepseek_v32, --enable-auto-tool-choice, --reasoning-parser deepseek_v3, --max-model-len 131072. Pin vLLM >= 0.11.2 for the multi-turn tool-call fix.
Claude Code via OPENAI_BASE_URL override, Roo Code as OpenAI Compatible, Cline same, Aider via --openai-api-base. SWE-bench-class scaffolds work out of the box because DeepSeek's training pipeline included these frameworks.
Open WebUI plus AnythingLLM live. Reasoning toggle exposed per conversation. Workspaces isolate sensitive document sets.
LangGraph workflows with context-management strategies (Summary, Discard-75) wired into checkpoint logic. Three workflows shipped: deep-research, repo-review, customer-triage.
OpenHands repo-autonomy. Expected SWE-bench Verified pass around 70 to 73 percent on routine tasks given the model's 72 to 74 percent benchmark range.
Evidence pack with tool-call success rates, throughput, reasoning-trace samples, runbook, and agent inventory delivered.
Fixed fee.
Hardware audit. Confirm GPU node or 512GB Mac. Run real reasoning prompts on V3.2 versus V3.2-Speciale. Pick variant by workload.
Serve via vLLM with FP8 on Hopper or Blackwell. Configure DeepSeek Sparse Attention for long-context cost savings. OpenAI-compatible API in front.
RAG with high-quality reranker (BGE reranker v2-m3). Tool-calling pipeline (V3.2 base only, Speciale does not support tools). RBAC, audit, cost telemetry per expert.
Pen-test cleared. MIT license NOTICE filed (one line). Production cutover with FP8 throughput baseline.
Yes. Both repository and model weights are MIT. No MAU cap, no commercial restriction, no attribution string in product. Counsel-friendly license that ships clean through any procurement review. You can redistribute, fine-tune, and embed in commercial products without contacting DeepSeek.
Mac Studio M3 Ultra 512GB at Q4 is the cheapest single-machine path. Around 8 to 15 tokens per second for one user. Fine for research, slow for production serving. Below 192GB unified memory you are forced into Q2 or 1.58-bit and the reasoning quality wobble shows up on hard tasks.
Base V3.2 supports tool calling and is the daily driver for agents, RAG, and general work. Speciale is the high-compute deep-reasoning variant, gold-medal class on IMO and IOI, but does NOT support tool calling. Run base for production. Add Speciale as a specialist endpoint for the hard reasoning queue if you have the GPU budget.
FP8 is near-lossless on H100 and H200. Q4_K_M loses 2 to 4 percent on commodity tasks. Q2_K shows 5 to 8 percent dips on math and structured reasoning. For a reasoning model, do not go below Q4 in production. Reasoning chains amplify quantization errors more than single-shot answers.
On agentic coding benchmarks DeepSeek V3.2 sits in the low-to-mid 70s on SWE-bench Verified (72-74 percent), a few points behind GPT-5.1-High and Gemini 3 Pro and roughly 7 to 9 points behind the leader, Claude Opus 4.5 at about 81 percent. On routine refactor work the gap is rarely visible. The decisive factor is licensing: DeepSeek V3.2 is MIT, runs on your hardware, no third-party processor, no per-token API spend. For HIPAA, GDPR, defense, or strict-NDA buyers, that benchmark gap is worth it. For frontier-quality output without the constraints, hybrid-route the hardest 10 percent of tasks to Claude via a private BAA endpoint.
--tokenizer-mode deepseek_v32 --tool-call-parser deepseek_v32 --enable-auto-tool-choice --reasoning-parser deepseek_v3. Without all four, tool calls return empty or reasoning leaks into content. Pin vLLM >= 0.11.2 for the multi-turn streaming fix (commit 36056). On Hopper prefer TP=2 with data and expert parallelism over TP=8 because of a kernel head-padding penalty in flashmla_sparse.py.
671B total, 37B active per token (MoE) Parameters
Fixed fee.
Day 1 you see the exact model, the exact hardware, the exact agent stack, the exact cost math, and the exact compliance posture.
ByKostadin StambolievFounder and Imaginator
Pineido, based in Strumica, North Macedonia, is a software company that builds websites, online shops, web and mobile apps, business automation and AI for companies at home and abroad. Work ships in 30-day sprints at a fixed price, and the client owns the code.
Published Updated