Deepseek Models

DeepSeek Models: V3.2, R1, Distills, and Production Caveats

GLM 5.2 is live! Opus-level intelligence at open-source rates. Pay per token on serverless. Try it today.

Blog

Deepseek Models The DeepSeek Model Lineup: V3.2, R1, and Distilled Variants Mapped to Production Workloads

PUBLISHED 2/27/2026

Table of Contents The DeepSeek Model Family: Five Variants, Five Deployment Decisions How to Pick the Right DeepSeek Variant for Your Workload Production Constraints: Tool Calling, Prompting, and Licensing What Breaks When You Self-Host DeepSeek How Fireworks Deploys DeepSeek: Serverless, On-Demand, and Enterprise How to Fine-Tune and Optimize DeepSeek on Fireworks Start Deploying DeepSeek on Fireworks

Table of Contents

Key Takeaways • DeepSeek's API alias remapping means deepseek-chat and deepseek-reasoner now both point to V3.2, so any team routing to those endpoints without pinning a version is hitting a different model than they think. • The five DeepSeek variants split across three distinct deployment decisions: V3.2 for general-purpose and agentic workloads, R1-0528 for deep reasoning with tool calls. The R1 distills handle cost-constrained reasoning, where an 8B model matches Qwen3-235B-thinking on reasoning benchmarks. • Tool use and thinking mode are mutually exclusive on V3.1. Few-shot prompting consistently degrades R1 performance. No variant exposes a clean API parameter to cap reasoning tokens. Design your pipeline around these constraints before you wire the integration. • Self-hosting DeepSeek surfaces documented failure modes including a 32k input token ceiling in certain Triton configurations and empty tool_calls arrays on distilled variants; we resolve these at the platform level on Fireworks On-Demand, which delivers ~250% better throughput and 50% lower latency than vLLM.

As most AI developers are well aware, DeepSeek has become one of the defining companies in the open-weights AI ecosystem. Founded in 2023, the Chinese lab made global headlines in January 2025 when the release of R1 triggered one of the largest single-day market sell-offs in recent memoletry — wiping billions from Nvidia, Broadcom, and ASML as investors confronted an uncomfortable reality: that a Chinese lab operating under strict GPU export controls had managed to train a frontier-competitive model with orders of magnitude less compute than anyone thought necessary. The panic reflected a long-held assumption — that China's lack of access to high-end chips meant its labs would perpetually lag behind Western frontier players like OpenAI and Anthropic — and R1 shattered it overnight. While markets eventually recovered, the impact on the AI industry has been lasting. DeepSeek remains the big blue whale of open-source AI — not just because of R1's benchmark performance, but because the lab has consistently pushed its most important architectural and training innovations directly into the open-source community. In doing so, it has done more than any other single organization to advance the idea that frontier AI capability can be unlocked through efficiency, not just scale. If you've been routing requests to deepseek-chat and assuming you know which model you're hitting, that assumption broke in late 2025. DeepSeek remapped both deepseek-chat and deepseek-reasoner to V3.2 ( DeepSeek API docs ), which means the lineup you thought you understood has shifted under you. The full family now spans five distinct release lines: DeepSeek-V3.2 (general-purpose), DeepSeek-R1 and DeepSeek-R1-0528 (reasoning-first), DeepSeek-V3.1 (hybrid reasoning), and six distilled variants from 1.5B to 70B parameters. Each carries distinct cost, latency, and licensing profiles. Picking the wrong one costs you in token budget or capability gaps that surface only after you're already in production. Dario Amodei argued that "DeepSeek-V3 was actually the real innovation" . The V3 engineering story backs that up: sparse attention and long-context efficiency, combined with agentic post-training, quietly made V3.x the more versatile production option while the discourse over-indexed on R1's reasoning capabilities. The rest of this guide covers what that means for each variant running on Fireworks AI, the inference platform built for production-grade model deployment.

The DeepSeek Model Family: Five Variants, Five Deployment Decisions

The DeepSeek ecosystem spans two architecture classes: MoE (Mixture of Experts) for the V3.x line and the full-scale R1 models, and dense transformer for the distilled variants that inherit R1's reasoning capability at smaller scale. The table below captures the lineup as it stands today. Variant Architecture Parameters Context Length Primary Use Case License DeepSeek-V3.2 MoE (sparse) 675.2B total 163.8k tokens General-purpose, agentic workflows, tool use Model License (see DeepSeek terms) DeepSeek-V3.1 MoE (sparse) 674.4B total 163.8k tokens Hybrid reasoning/direct-answer, tool use Model License (see DeepSeek terms) DeepSeek-R1 MoE (V3 base) 671B total 163.8k tokens Deep reasoning, math, code, long chain-of-thought MIT DeepSeek-R1-0528 MoE (V3 base) 674.4B total 163.8k tokens Deep reasoning with function calling MIT R1 Distills (×6) Dense (Qwen/Llama base) 1.5B–70B (includes 7B, 8B, 14B, 32B, 70B variants) Varies (~131k typical) Cost-constrained reasoning, edge deployment Apache 2.0 (Qwen-based); Llama license (Llama-based)

Several details in this table matter more than they appear. The API alias remapping means deepseek-chat now points to V3.2, not V3 ( DeepSeek API docs ). DeepSeek-V3.1 occupies a hybrid position: it switches between direct-answer and chain-of-thought reasoning by changing the chat template, with no separate model endpoint required. DeepSeek-R1-0528 added function calling support that DeepSeek-R1 lacked, which reopens the agentic deployment question for reasoning-first workloads. The six R1 distills inherit their licenses from their base models. Qwen-based distills (1.5B, 7B, 14B, 32B) carry Apache 2.0. Llama-based distills (8B, 70B) carry the Llama license , which includes a 700-million monthly active user threshold for commercial use. These are different production constraints that do not collapse into a single "MIT" label. The architecture split also has direct cost implications. V3.x's MoE design activates only a subset of parameters per token, keeping inference cost low despite the 675B total parameter count. The dense...