DeepSeek V3 Review 2026: Benchmarks, Legacy Status & Migration

Should you still build on DeepSeek V3 in October 2026, or has the lineage moved on? This DeepSeek V3 review answers that. V3 was a landmark December 2024 release: DeepSeek’s technical report describes a 671B-parameter Mixture-of-Experts model, with 37B parameters active per token, as comparable to leading closed-source models such as GPT-4o and Claude-3.5-Sonnet. Since then the API has moved through V3-0324, V3.1, V3.2 and the V4 Preview (24 April 2026) to today’s DeepSeek-V4.1-Flash-1017 and DeepSeek-V4-Pro-0813. This review summarises DeepSeek’s published evidence on V3: benchmarks, pricing history, strengths, weaknesses and what to use instead today.

Our verdict: a milestone model, now superseded

Our assessment, based on DeepSeek’s V3 technical report (December 2024), the V3 model card and DeepSeek’s API documentation: V3 was a milestone open-weight release, which DeepSeek’s report called “the strongest open-source model currently available” at the time. The V3 weights remain downloadable on Hugging Face under the DeepSeek Model License (code MIT). As of 22 October 2026, however, V3 is not offered on DeepSeek’s hosted API; the current models are deepseek-flash (DeepSeek-V4.1-Flash-1017) and deepseek-v4-pro (DeepSeek-V4-Pro-0813).

The headline: in our assessment, V3 still makes sense mainly if you already self-host its open weights for a steady workload, or need the December 2024 checkpoint for research. For a hosted API model today, start with deepseek-flash and compare the lineup in the DeepSeek models hub.

Who should use it, who shouldn’t

Use V3 if:

• You already run a self-hosted V3 deployment and the workload is stable.
• You need the December 2024 checkpoint itself, for example for reproducibility research.
• You are studying MoE architecture: the report documents V3’s auxiliary-loss-free load balancing and multi-token prediction objective in detail.

Don’t use V3 if:

• You are building new on the hosted API. V3 is no longer offered there; until 24 July 2026 (15:59 UTC) the legacy deepseek-chat and deepseek-reasoner IDs routed to DeepSeek V4-Flash (non-thinking / thinking), and they have since been retired.
• You need long context. V3’s weights support 128K tokens (model card); today’s deepseek-flash and deepseek-v4-pro offer a 1M-token context and up to 384K output tokens (pricing page, 22 October 2026).
• You need built-in thinking mode. V3 was non-thinking; reasoning lived in the separate R1 model, released on 20 January 2025.

How we assessed it

DS Guide Editorial does not run its own evaluations. This review draws on DeepSeek’s official documentation, technical report, pricing and published benchmark tables, so every benchmark figure here is DeepSeek’s own, labelled as vendor-reported. Sources:

• DeepSeek-V3 technical report, arXiv 2412.19437 (v1 27 Dec 2024; v2 18 Feb 2025), Tables 1, 6 and 7.
• The DeepSeek-V3 model card and the licence tags of later DeepSeek checkpoints on Hugging Face (checked 22 Oct 2026).
• DeepSeek release notes for V3 (26 Dec 2024), R1 (20 Jan 2025), V3.1 (21 Aug 2025), the V4 Preview (24 Apr 2026) and V4.1-Flash (10 Sep 2026).
• The DeepSeek API Change Log, entries of 26 Dec 2024, 24 Mar 2025, 21 Aug 2025, 1 Dec 2025, 24 Apr 2026, 13 Aug 2026, 10 Sep 2026 and 17 Oct 2026.
• The Models & Pricing page, Thinking Mode guide, Chat Completions API reference and Multi-round Conversation guide (checked 22 Oct 2026).
• DeepSeek’s privacy policy (last updated 10 Feb 2026; checked 22 Oct 2026), the OpenAI, Anthropic and Google pricing pages, and OpenAI’s deprecations page (as of 22 October 2026).

Architecture (per DeepSeek’s technical report)

DeepSeek’s V3 technical report describes a Mixture-of-Experts model with 671B total parameters, of which 37B are activated for each token. It uses Multi-head Latent Attention (MLA) and the DeepSeekMoE architecture, and introduces an auxiliary-loss-free load-balancing strategy. DeepSeek pre-trained V3 on 14.8 trillion tokens, then applied supervised fine-tuning and reinforcement learning. The report presents MLA and DeepSeekMoE as the basis for efficient inference and cost-effective training.

What DeepSeek’s published results show by task type

Scores are DeepSeek’s own, from the V3 report’s Table 6 (chat models, December 2024) unless stated; the report says closed models were evaluated through their APIs.

Coding

DeepSeek’s report lists 82.6 on HumanEval-Mul (a multilingual set covering eight programming languages), the highest of the models it compared: Claude-3.5-Sonnet-1022 scored 81.7, GPT-4o-0513 80.5, Qwen2.5-72B-Instruct 77.3 and LLaMA-3.1-405B-Instruct 77.2. On SWE-bench Verified, run with the agentless framework, V3 resolved 42.0% of tasks against Claude’s 50.8%, and the report says V3 “trails behind Claude-Sonnet-3.5-1022” on engineering tasks. The Change Log entry of 24 March 2025 says V3-0324 raised LiveCodeBench from 39.2 to 49.2 and improved front-end code-generation accuracy.

Mathematical reasoning

V3 posted the top Table 6 score on all three math benchmarks: MATH-500 90.2, AIME 2024 39.2 and CNMO 2024 43.2. The report also claims V3 “even outperforms o1-preview on specific benchmarks, such as MATH-500” (o1-preview is not in Table 6). For step-by-step reasoning, DeepSeek’s dedicated model at the time was R1 (released 20 January 2025 under the MIT licence) — see DeepSeek R1.

Writing

The report does not score writing style directly. Its closest measures are the open-ended evaluations in Table 7, judged by GPT-4-Turbo-1106: Arena-Hard 85.5 and AlpacaEval 2.0 70.0 (length-controlled win rate), against 85.2 and 52.0 for Claude-Sonnet-3.5-1022. The report also notes, from an ablation on DeepSeek-V2.5, that distilling from R1 “substantially increases the average response length”, a trade-off DeepSeek says it balanced in V3’s settings. For prompting today’s deepseek-flash and deepseek-v4-pro, see the DeepSeek prompt engineering tutorial.

Benchmark snapshot

The original V3 (chat) figures DeepSeek published, beside the two closed models in the same table:

DeepSeek’s report says V3 outperformed all other open-source models on MMLU, MMLU-Pro and GPQA, though its own Table 6 lists LLaMA-3.1-405B-Instruct at 88.6 on MMLU, a shade above V3’s 88.5. Source: DeepSeek-V3 technical report, arXiv 2412.19437, Table 6 (v2, 18 Feb 2025).

Value for money

DeepSeek’s 26 December 2024 release note kept V2 pricing until 8 February 2025, then set V3 list prices of $0.27 per million input tokens on a cache miss, $0.07 on a cache hit and $1.10 per million output tokens. Those rates are no longer current. DeepSeek’s V3.1 release note (21 August 2025) introduced new pricing from 16:00 UTC on 5 September 2025, when the off-peak discounts of that era also ended. Today’s separate peak/off-peak scheme dates from 16 August 2026: peak hours are 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday, excluding Chinese public holidays, and all other hours are off-peak at half the peak rate. For today’s DeepSeek API pricing, do not quote the V3 numbers as live.

Worked example (hypothetical): deepseek-flash today versus V3’s 2025 list prices

As a hypothetical illustration, take 1,000,000 calls per month, each with a 2,000-token cached system prompt, a 200-token uncached user message and a 300-token response, at list prices as of 22 October 2026 (peak and off-peak) and at V3’s 2025 list prices:

deepseek-flash, peak / off-peak (22 October 2026)
Cache hit  : 2,000,000,000 ÷ 1,000,000 × $0.006 / $0.003 = $12.00 / $6.00
Cache miss :   200,000,000 ÷ 1,000,000 × $0.30 / $0.15   = $60.00 / $30.00
Output     :   300,000,000 ÷ 1,000,000 × $1.20 / $0.60   = $360.00 / $180.00
Total                                                      $432.00 / $216.00

V3 standard list prices from 8 Feb 2025 (no longer offered)
Cache hit  : 2,000,000,000 ÷ 1,000,000 × $0.07 = $140.00
Cache miss :   200,000,000 ÷ 1,000,000 × $0.27 = $54.00
Output     :   300,000,000 ÷ 1,000,000 × $1.10 = $330.00
Total                                            $524.00

deepseek-v4-pro, peak / off-peak (22 October 2026)
Cache hit  : 2,000,000,000 ÷ 1,000,000 × $0.044 / $0.022 = $88.00 / $44.00
Cache miss :   200,000,000 ÷ 1,000,000 × $1.32 / $0.66   = $264.00 / $132.00
Output     :   300,000,000 ÷ 1,000,000 × $3.96 / $1.98   = $1,188.00 / $594.00
Total                                                      $1,540.00 / $770.00

All lines are exact to the cent, so no rounding was needed; only the ratios are rounded. V3’s 2025 list prices, excluding that era’s off-peak discounts, come to about 1.2 times today’s all-peak deepseek-flash bill ($524.00 ÷ $432.00 ≈ 1.21) and about 2.4 times the all-off-peak bill ($524.00 ÷ $216.00 ≈ 2.43). The saving comes from far cheaper cache hits ($0.006 peak / $0.003 off-peak against $0.07) and off-peak scheduling; at peak, cache-miss input ($0.30) and output ($1.20) now cost more than V3’s $0.27 and $1.10. deepseek-flash is the cheaper of DeepSeek’s two current models for this profile, and running it off-peak halves the bill; deepseek-v4-pro costs about 3.6 times as much in either period ($1,540.00 ÷ $432.00 ≈ 3.56). Use the DeepSeek pricing calculator for your own numbers.

How V3 actually worked through the API

For a V3-era integration, two points matter. First, the chat API is stateless: requests go to the OpenAI-compatible POST /chat/completions endpoint, and the client resends the full message history on every call (Multi-round Conversation guide, checked 22 October 2026). Second, V3 was reached through the deepseek-chat alias, which served V3 from 26 December 2024 until it moved to V3-0324 on 24 March 2025; after later upgrades it routed to V4-Flash (non-thinking) until it was retired at 15:59 UTC on 24 July 2026.

A minimal Python call to today’s endpoint (POST https://api.deepseek.com/chat/completions) with deepseek-flash and thinking switched off, matching the non-thinking V3-era deepseek-chat:

from openai import OpenAI

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

resp = client.chat.completions.create(
    model="deepseek-flash",
    messages=[{"role": "user", "content": "Summarise this PR diff."}],
    temperature=0.0,
    max_tokens=1024,
    extra_body={"thinking": {"type": "disabled"}},
)
print(resp.choices[0].message.content)

temperature=0.0 takes effect only because thinking is disabled: thinking is the default (effort high), and in thinking mode temperature is accepted but has no effect (Thinking Mode guide, checked 22 October 2026). To use thinking, drop temperature and pass reasoning_effort="high" with extra_body={"thinking": {"type": "enabled"}}.

Strengths that aged well

• Architectural openness. A detailed report plus published weights for a 671B-parameter MoE with 37B activated per token give researchers a documented reference design.
• Training efficiency. DeepSeek’s report puts V3’s full training at 2.788M H800 GPU hours, or $5.576M at an assumed $2 per GPU-hour, excluding prior research and ablations (Table 1).
• OpenAI SDK compatibility. V3 arrived behind the existing deepseek-chat name with the API unchanged (Change Log, 26 December 2024); today’s API also accepts the Anthropic format.
• Open weights on Hugging Face. The V3 checkpoint is still downloadable (checked 22 October 2026) under the DeepSeek Model License, which allows commercial use, with the code under MIT; the V3-0324, V3.1, V3.2, R1, V4-Pro, V4-Flash, V4.1-Flash and V4.1-Flash-1017 checkpoints are tagged MIT.

Documented weaknesses

• 128K context ceiling. V3’s weights top out at 128K tokens (model card), against 1M for today’s deepseek-flash and deepseek-v4-pro.
• No native thinking mode. Reasoning required the separate R1 model. Since V3.1 (21 August 2025) one model serves both modes; today’s models default to thinking (effort high), which extra_body={"thinking": {"type": "disabled"}} switches off and reasoning_effort (low / high / max) tunes — see the DeepSeek API documentation.
• Large deployment unit. The report’s limitations section says “the recommended deployment unit for DeepSeek-V3 is relatively large, which might pose a burden for small-sized teams”.
• Function calling and factual recall. The 24 March 2025 Change Log credits V3-0324 with “fixing issues from previous V3 versions” in function calling, and Table 6 puts V3 behind both closed models on SimpleQA.
• Hosted-service privacy. DeepSeek’s privacy policy (last updated 10 February 2026, checked 22 October 2026) states: “To provide you with our services, we directly collect, process and store your Personal Data in People’s Republic of China.” Self-hosted weights avoid this; see DeepSeek privacy.

Competitor context

DeepSeek’s V3 report (December 2024) compared the chat model with GPT-4o-0513 and Claude-3.5-Sonnet-1022, alongside Qwen2.5-72B-Instruct and LLaMA-3.1-405B-Instruct, and lists V3’s HumanEval-Mul score as 82.6 against 80.5 and 81.7 for the two closed models (vendor-reported). For head-to-heads, see DeepSeek V3 vs GPT-4o and DeepSeek vs Claude.

Those reference points have aged. Standard prices as of 22 October 2026 (input / cached input / output per 1M tokens): GPT-4o (the gpt-4o model, not the 0513 snapshot in DeepSeek’s table) $2.50 / $1.25 / $10.00 on OpenAI’s pricing page, though OpenAI has announced its shutdown for 14 January 2027; GPT-5.5 $5.00 / $0.50 / $30.00 for prompts up to 272K tokens (same page); Claude Opus 4.8 $5.00 / $0.50 / $25.00, now a legacy model, on Anthropic’s Claude pricing page; and Gemini 3.1 Pro Preview $2.00 / $0.20 / $12.00 for prompts up to 200K tokens on Google’s Gemini API pricing page. Newer models now head each vendor’s lineup.

Migration path: from V3-era IDs to today’s models

1. Replace model="deepseek-chat" (retired 24 July 2026) with model="deepseek-flash" and pass extra_body={"thinking": {"type": "disabled"}} to keep non-thinking behaviour; base_url="https://api.deepseek.com" stays the same.
2. For former deepseek-reasoner workloads, use deepseek-flash (or deepseek-v4-pro, which lacks Flash’s vision support) with reasoning_effort="high" and extra_body={"thinking": {"type": "enabled"}}, which match the default; reasoning_content comes back alongside content. (DeepSeek’s 10 September 2026 note claims V4.1-Flash benchmarks ahead of V4-Pro.)
3. Prompts sized for V3’s 128K context fit today’s 1M-token window, so chunked workflows can be consolidated.
4. Re-run your eval set. To tune temperature, disable thinking first.
5. deepseek-chat and deepseek-reasoner were retired at 15:59 UTC on 24 July 2026. The later legacy names deepseek-v4-flash and deepseek-v4-flash-vision-exp are still accepted but retired, served by DeepSeek-V4.1-Flash-1017 at the Flash price; new code should use deepseek-flash.

Verdict

Our assessment, based on DeepSeek’s technical report (December 2024) and official Change Log: V3 was an open-weight model whose vendor-reported results edged GPT-4o-0513 and Claude-3.5-Sonnet-1022 on MMLU and beat both on several math and algorithmic-coding benchmarks, while trailing Claude on SWE-bench Verified, GPQA-Diamond and MMLU-Pro. As engineering and as a research artefact, it remains worth studying. As of 22 October 2026 it is not offered on DeepSeek’s hosted API, so build on deepseek-flash (DeepSeek-V4.1-Flash-1017) or deepseek-v4-pro (DeepSeek-V4-Pro-0813). Browse DeepSeek reviews for the rest of the lineage.

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.