Fine-Tune DeepSeek V4 with Ertas

Overview

DeepSeek V4, released April 24, 2026, is the largest and most capable open-weight model available at the time of release. The flagship V4 Pro variant uses a 1.6 trillion parameter mixture-of-experts architecture with approximately 49B parameters active per token, paired with a 1 million token context window. A smaller V4 Flash variant ships alongside it at 284B total / 13B active parameters, also with 1M context, targeting deployment scenarios where the Pro model's memory footprint is impractical.

The V4 release continues the architectural innovation that made DeepSeek's prior generation a defining moment in open-source AI. V4 builds on the DeepSeek Sparse Attention (DSA) mechanism introduced in V3.2, refines the MoE expert routing topology, and applies a substantially expanded reinforcement learning post-training pipeline. The cumulative effect is a model that, on the BenchLM aggregate intelligence index (87) at the time of release, leads all open-weight models and significantly narrows the gap with frontier proprietary systems like GPT-5.5 and Claude Opus 4.7.

Unlike DeepSeek-R1, V4 is not a dedicated reasoning-only model. Instead, V4 incorporates a thinking-mode toggle similar to Qwen 3+: the same checkpoint serves both direct-response (chat) and extended-reasoning (reasoner) modes via a control flag at inference time. This unification reduces operational complexity for production deployments compared to maintaining separate R1-style reasoning models alongside V3-style instruction-tuned models.

V4 is released under the DeepSeek License — a permissive MIT-style commercial license. The model weights are available on Hugging Face under `deepseek-ai/DeepSeek-V4-Pro` and `deepseek-ai/DeepSeek-V4-Flash`, with corresponding `-Base` variants for further fine-tuning. The license terms permit broad commercial use including model serving, derivative training, and proprietary integration.

Key Features

The 1M token context window is one of the largest publicly deployed in any open-weight model. Combined with the DSA sparse attention mechanism, V4 maintains usable performance on long-context retrieval and reasoning tasks far better than naive RoPE-extended models. While effective context (the range over which the model retains >90% retrieval accuracy) is smaller than the advertised 1M, the model is genuinely usable for full-codebase analysis, long-document QA, and multi-document synthesis at scales no previous open-weight release could handle.

DeepSeek Sparse Attention reduces the quadratic compute cost of long-context attention by routing each query token to a learned subset of key tokens rather than attending to all of them. This delivers the dual benefit of supporting much longer contexts than dense attention would allow on equivalent hardware, while also reducing inference cost on shorter sequences as compared to a dense attention baseline at the same model scale.

The unified thinking mode is operationally significant. Production deployments can dispatch most queries directly through fast non-thinking inference, then escalate hard queries to reasoning mode by passing a single control parameter — without swapping model weights or routing across separate endpoints. This pattern significantly simplifies the operational topology of agentic systems compared to the prior generation, where R1 and V3 were two distinct deployments.

V4 also continues DeepSeek's strong performance on coding (SWE-Bench Verified ~73%), reasoning (AIME 2025 in the high 70s%), and math benchmarks while improving on multilingual capabilities and tool-use fidelity. The model is one of the strongest open-weight choices for tool-using agents requiring high reliability on function-call schemas.

Fine-Tuning with Ertas

DeepSeek V4's scale makes full fine-tuning impractical for most teams, but Ertas Studio supports QLoRA fine-tuning on V4 Flash, the 284B/13B variant, on multi-GPU server setups (8x A100 80GB or equivalent). V4 Flash QLoRA at 4-bit base quantization plus LoRA adapters on attention and MoE expert projections requires approximately 280-340GB of total VRAM at typical sequence lengths, distributed across the GPU set with tensor parallelism.

For most teams interested in DeepSeek V4 capability without the multi-GPU footprint, Ertas Studio's recommended approach is to fine-tune one of the DeepSeek-R1 distilled variants (Qwen 7B, 14B, 32B, or Llama 70B distilled) using V4 Pro as a teacher model for synthetic data generation. This approach delivers the V4 reasoning style at the deployment cost of a dense model in the 7B-70B range — tractable on a single GPU and far cheaper to serve.

After fine-tuning, Ertas Studio exports to GGUF format. V4 Flash quantized to Q4_K_M is approximately 145GB, requiring a multi-GPU server or large-memory CPU inference host (256GB+ RAM). Distilled fine-tuned models export at standard sizes for their base parameter counts and deploy cleanly on Ollama, llama.cpp, or vLLM. For teams running V4 Pro as a teacher and a smaller distilled student in production, Ertas Studio supports the full pipeline including synthetic data generation, distillation training, and final quantization.

Use Cases

V4 Pro's 1M context window unlocks use cases that were previously infeasible on open-weight infrastructure: full-codebase code review where the model considers all source files simultaneously, long-document legal or financial analysis where the entire contract or filing fits in a single prompt, and multi-document synthesis tasks like literature reviews or competitive intelligence where dozens of sources must be reasoned over jointly.

V4 Flash is the more practical choice for general-purpose production serving. With 13B active parameters, it serves at competitive token-per-second rates while delivering quality that approaches V4 Pro on standard benchmarks. The 1M context is preserved, making Flash an ideal choice for RAG systems with very large retrieval result sets.

The unified thinking mode makes V4 a strong fit for agentic systems that need adaptive reasoning depth. Customer support agents can run primarily in fast direct-response mode, escalating to reasoning mode only for genuinely complex tickets. Coding agents can use direct mode for simple completions and reasoning mode for architectural decisions or debugging. This pattern significantly reduces inference cost compared to running pure reasoning-mode inference uniformly.

Hardware Requirements

V4 Pro at Q4_K_M quantization requires approximately 820GB of total memory, which in practice means an 8x H100 80GB or 8x A100 80GB server, or a CPU inference host with 1TB+ of RAM. Active parameter count of 49B determines generation throughput, so once loaded the model serves at speeds comparable to a 49B dense model. This is large-server territory, not consumer or single-workstation deployment.

V4 Flash at Q4_K_M is approximately 145GB. This fits on a 4x A100 80GB or 2x H100 with margin, or a CPU host with 256GB+ RAM. Active parameters of 13B mean inference speed is comparable to a 13B dense model — well-suited for high-throughput API serving with reasonable per-request latency. For teams that want DeepSeek V4 quality without the V4 Pro hardware footprint, Flash is the practical recommendation.

For fine-tuning in Ertas Studio: V4 Flash QLoRA needs approximately 280-340GB total VRAM (multi-GPU server). V4 Pro QLoRA is impractical for most teams — the recommended approach is distillation onto a smaller base model. Distilled R1-style fine-tuning of Qwen 32B or Llama 70B in Ertas Studio requires the standard 20-48GB VRAM for those base models with QLoRA.