If you have landed here trying to work out whether DeepSeek Coder is still worth running in 2026 — and which version you should actually pull from Hugging Face — the short answer is that the original Coder line has been superseded by DeepSeek Coder V2, and V2 has in turn been largely folded into the general-purpose V4 models that now handle coding workloads on the API. The original weights still work, still hold up on certain benchmarks, and still run on a single GPU at the smaller sizes. But the centre of gravity has moved. This guide walks through what each release actually shipped, the benchmark numbers from the original papers, how the lineage connects to today’s V4 API, and where DeepSeek Coder makes sense in a 2026 workflow.

What is DeepSeek Coder?

DeepSeek Coder is an open-weight family of code-specialised language models from DeepSeek, first released in late 2023. The line comprises a series of code language models trained from scratch on both 87% code and 13% natural language in English and Chinese, with each model pre-trained on 2T tokens, ranging from 1B to 33B versions. A second generation — DeepSeek Coder V2 — landed in June 2024 and shifted the architecture to Mixture-of-Experts (MoE). Since then, the coding capabilities have been merged into DeepSeek’s general-purpose flagship models, with DeepSeek V4 now serving the strongest coding benchmarks on the hosted API.

Architecture and lineage

There are two distinct generations under the DeepSeek Coder name, plus a successor path that runs through the generalist models.

Generation 1: DeepSeek Coder (November 2023)

The original release was a dense decoder-only Transformer family. The series comprises open-source code models varying in size from 1.3B to 33B, including base and instructed versions for each size. Each model was trained from scratch on 2 trillion tokens sourced from 87 programming languages, with pre-training data organised at the repository level to enhance understanding of cross-file context within a repository. In addition to next-token prediction loss, training incorporated the Fill-In-Middle (FIM) approach. The context window was 16K tokens.

Generation 2: DeepSeek Coder V2 (June 2024)

V2 was a complete re-architecture on the DeepSeek-MoE backbone. DeepSeek released Coder V2 with 16B and 236B total parameters based on the DeepSeekMoE framework, which has active parameters of only 2.4B and 21B, including base and instruct models. The pre-training dataset was 60% source code, 10% math corpus, and 30% natural language. The source code consisted of 1,170B code-related tokens from GitHub and CommonCrawl. The corpus expanded from 86 to 338 programming languages compared to the original DeepSeek Coder. Context extended to 128K, allowing processing of large code repositories and complex problems.

What happened next

DeepSeek did not release a “Coder V3.” Instead, coding capability was absorbed into the generalist reasoning models — DeepSeek V3, DeepSeek R1, V3.2, and now the V4 family. If you want the strongest DeepSeek coding performance on the API today, you call deepseek-v4-pro or deepseek-v4-flash, not a dedicated Coder endpoint.

Benchmarks from the primary sources

Numbers below are taken directly from the respective technical reports. These are historical benchmarks from the original evaluation context; versions of competing models (GPT-4-Turbo, Claude 3 Opus, GPT-3.5-Turbo) are named explicitly to avoid cross-generation confusion.

DeepSeek Coder (Gen 1) — base models

DeepSeek-Coder-Base achieved strong benchmark performance at release, with an average accuracy of 50.3% on HumanEval and 66.0% on MBPP. Compared to the similarly sized open-source model CodeLlama-Base 34B, it shows improvements of 9% and 11% in accuracy, respectively. DeepSeek-Coder-Instruct 33B was the only open-sourced model at the time that outperformed OpenAI’s GPT-3.5-Turbo on LeetCode-style tasks, though a substantial gap remained with GPT-4-Turbo.

DeepSeek Coder V2 — 236B MoE

DeepSeek Coder V2 achieved a 90.2% score on HumanEval, a 76.2% score on MBPP+ (setting a new high mark on the EvalPlus evaluation pipeline), and a 43.4% score on LiveCodeBench (Jain et al., 2024) on questions from December 2023 to June 2024. Additionally, DeepSeek-Coder-V2 was the first open-source model to surpass a score of 10% on SWE-Bench. For current SWE-Bench Verified numbers on V4-Pro (80.6% in DeepSeek’s V4 announcement), see the DeepSeek benchmarks 2026 roundup.

Strengths

• Repo-level training. Both generations organised pre-training at the repository level, not the file level, which helps the model reason across imports and shared types. The original used a 16K window; V2 extended to 128K.
• Fill-in-the-Middle is native. FIM has been a first-class pre-training objective from day one. This matters for editor integrations where the model has to complete a span inside an existing buffer, not just continue from a cursor position. If you are wiring a completion into an editor, see the DeepSeek with VS Code walkthrough.
• Small-model efficiency. The 6.7B and Lite-16B tiers punch well above their weight. DeepSeek-Coder-V2-Lite-Base, despite having only 2.4 billion active parameters, achieves code completion capabilities in Python comparable to the DeepSeek-Coder-Base 33B model.
• Commercial use permitted. The code repository is licensed under the MIT License. Use of DeepSeek Coder models is subject to the Model License. DeepSeek Coder supports commercial use. Note that the weights sit under a separate DeepSeek Model License for the original Coder and Coder V2 — not MIT — so read the LICENSE file before shipping a derivative.

Weaknesses

• It is a 2023/2024 model. Training cut-off predates a large fraction of modern framework versions. For code using recent APIs, a current frontier model will generally do better.
• No reasoning mode. Coder V2 is not a reasoning model. It provides direct responses without extended chain-of-thought reasoning. For complex multi-file refactors, a reasoning-capable model returning reasoning_content alongside the final content tends to produce better plans.
• Large variants are hardware-heavy. To use DeepSeek-Coder-V2 in BF16 format for inference, 80GB×8 GPUs are required. That is 640 GB of VRAM for the 236B model — not a homelab deployment.
• Benchmarks trail V4. For any new production coding workload, the V4 family’s current SWE-Bench Verified numbers make the case for using the generalist API instead.

How to access DeepSeek Coder

Open weights on Hugging Face

All original Coder and Coder V2 weights remain live on the DeepSeek organisation on Hugging Face. The practical picks:

• deepseek-ai/deepseek-coder-6.7b-instruct — runs on a single consumer GPU with quantisation; reasonable FIM.
• deepseek-ai/deepseek-coder-33b-instruct — best Gen-1 quality; ~66 GB of VRAM in BF16.
• deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct — 16B total / 2.4B active; good throughput-per-dollar on a single 80 GB GPU.
• deepseek-ai/DeepSeek-Coder-V2-Instruct — 236B MoE; multi-GPU only.

For a step-by-step local deployment walkthrough, see how to install DeepSeek locally or the running DeepSeek on Ollama guide for the quantised route.

Hosted API — current model IDs

DeepSeek does not expose the Coder models as distinct API endpoints today. Chat requests hit POST /chat/completions, the OpenAI-compatible endpoint, at https://api.deepseek.com, and the model field takes one of:

• deepseek-v4-pro — 1.6T total / 49B active, frontier tier (released April 24, 2026).
• deepseek-v4-flash — 284B total / 13B active, cost-efficient tier.
• deepseek-chat and deepseek-reasoner — legacy IDs that currently route to deepseek-v4-flash (non-thinking and thinking modes, respectively). These IDs retire on 2026-07-24 at 15:59 UTC; migrating is a one-line model= swap.

A minimal Python call using the OpenAI SDK, demonstrating reasoning_effort and a coding-appropriate temperature of 0.0:

from openai import OpenAI

client = OpenAI(
    base_url="https://api.deepseek.com",
    api_key="YOUR_KEY",
)

resp = client.chat.completions.create(
    model="deepseek-v4-pro",
    messages=[
        {"role": "system", "content": "You write idiomatic Python."},
        {"role": "user", "content": "Refactor this module for async IO: ..."},
    ],
    temperature=0.0,
    max_tokens=8000,
    reasoning_effort="high",
    extra_body={"thinking": {"type": "enabled"}},
)

print(resp.choices[0].message.reasoning_content)
print(resp.choices[0].message.content)

The API is stateless — your client must resend the conversation history on every request. Unlike the web chat and app, nothing is remembered server-side between calls. Both V4 tiers default to a 1,000,000-token context window with output up to 384,000 tokens. FIM completion (Beta) and Chat Prefix Completion (Beta) are available in non-thinking mode. For the full surface, see the DeepSeek API documentation.

Pricing snapshot

Coder-specific pricing tiers no longer exist. Coding workloads pay the standard V4 rates. As of April 2026, per 1M tokens:

Worked example for a code-review agent at V4-Flash rates, 1,000,000 calls per month with a 2,000-token cached system prompt, a 200-token diff, and a 300-token review:

• Cached input: 2,000 × 1,000,000 = 2,000,000,000 tokens × $0.028/M = $56.00
• Uncached input: 200 × 1,000,000 = 200,000,000 tokens × $0.14/M = $28.00
• Output: 300 × 1,000,000 = 300,000,000 tokens × $0.28/M = $84.00
• Total: $168.00 per month

Always cost the uncached user message separately — the system-prompt cache hit does not cover the per-call diff. Off-peak discounts ended on 2025-09-05 and have not returned. Confirm current rates on the official DeepSeek pricing page before committing a budget; you can also model scenarios with the DeepSeek pricing calculator.

Best use cases

• Local code completion on commodity hardware — Coder-V2-Lite or the 6.7B original is still a solid pick when an internet round-trip is unacceptable. See DeepSeek for coding for workflow patterns.
• On-premise compliance workloads — enterprises that cannot send code off-site run the open weights behind their firewall.
• Fine-tuning a base coder — Coder V2-Lite-Base is a practical starting point for domain-specific fine-tunes. See the fine-tuning guide.
• IDE assistants and agentic coders — via API, target V4-Pro for hard tasks and V4-Flash for autocomplete.

Alternatives worth comparing

If you are deciding between DeepSeek Coder and a commercial assistant, the head-to-head at DeepSeek Coder vs Copilot covers latency, privacy, and completion quality. For a broader open-source survey, see DeepSeek alternatives for coding. To place the Coder line in the full DeepSeek family, browse the DeepSeek models hub.

Verdict

DeepSeek Coder still earns a slot in 2026, but a narrower one than in 2024. For local inference, fine-tuning, or any workload that cannot leave your network, the open weights — particularly Coder V2-Lite — are among the stronger permissively licensed code models available. For hosted coding on the API, the correct answer is now V4-Flash for routine work and V4-Pro when the SWE-Bench lift pays for itself.

Last verified: 2026-04-24. DeepSeek AI Guide is an independent resource and is not affiliated with DeepSeek or its parent company. Model IDs, pricing and API behaviour change; check the official DeepSeek documentation and pricing page before committing to a production decision.