DeepSeek V4 Pro

Type: AI model

DeepSeek V4 Pro is the high-capability variant of DeepSeek’s V4 series previewed April 2026, a mixture-of-experts model with about 1.6T total and 49B active parameters and a 1-million-token context window. It is positioned near the frontier at low cost with open weights. GROUNDING tracks DeepSeek V4 Pro’s benchmarks, pricing, and availability.

Recent Updates

• 2026-06-28: A Solo Developer’s GGUF Coding and Agentic Gemma Models Hit Hugging Face Trending (量子位) · qbitai.com — Code Agents Agents Tool Use Context Engineering LLM Evals Hugging Face · Zhipu · Baidu · Qwen · NVIDIA · Microsoft · Krea · Cursor · llama.cpp · Ollama · LM Studio Jan · Gemma yuxinlu1 逯雨鑫 听雨 · GLM-5.2 · Unlimited-OCR AgentWorld LocateAnything FastContext · MiniMax M3 · Kimi-K2.7-Code Krea-2-Turbo Krea-2-Raw · Fable 5 · Gemma 4 12B · Composer 2.5 · Claude Opus 4.8 gemma-4-12B-it-Claude-4.6-4.8-Opus-GGUF Mellum2 · Qwen3.6 · Qwen3.6-27B
• 2026-06-29: Paper proposes a signal-coverage matrix for autoformalization evaluation (cs.CL updates on arXiv.org) · arxiv.org — LLM Evals DeepSeek
• 2026-06-29: CEO-Bench suggests AI still struggles as a standalone company operator (量子位) · qbitai.com — Agents Tool Use Code Agents LLM Evals Context Engineering Princeton University Anthropic · Google · DeepSeek · OpenAI · QbitAI Jay Steve Jobs Jensen Huang Ilya Sutskever · GLM-5.1 · Claude Haiku 4.5 · Gemini 3 Flash Grok 4.20 · Claude Fable 5 · Claude Opus 4.8 · GPT 5.5 · Claude Opus 4.7 · Qwen 3.7 Max · GLM-5.2 · kimi-k2.6
• 2026-06-29: DeepSeek V4 is slated for mid-July with peak-hour API pricing (Hacker News) · kucoin.com — DeepSeek ME News BlockBeats KuCoinFlash · Hacker News · DeepSeek V4 · DeepSeek V4 Flash
• 2026-06-29: ClinePass adds access to latest open-weight models without API key juggling (DAIR.AI) · twitter.com — Open Source LLMs DAIR.AI · Cline · DeepSeek · MiniMax · GLM-5.2 · Kimi-K2.7-Code Mimo 2.5 · MiniMax M3
• 2026-06-30: Meituan says LongCat 2.0 was trained on Chinese chips and large-scale long-context data (‌эйай ньюз) · arxiv.org — Long Context Embeddings Meituan · Huawei · NVIDIA · Google · OpenRouter · LongCat 2.0 LongCat Flash-Lite Owl Alpha Huawei Ascend 910C
• 2026-06-30: Latent Space recap highlights Cursor remote agents, open-weight access, and Meta Brain2Qwerty v2 (Latent.Space) · latent.space — Agents Code Agents Tool Use LLM Evals Open Source LLMs Latent Space AIEWF · Meta · Cursor · Cline · Cognition · Arena · DeepSeek · Qwen · MiniMax JeanRemiKing kimmonismus Garry Tan stalkermustang teortaxesTex Brain2Qwerty v2 · DeepSeek V4 Flash · Qwen3-4B
• 2026-07-01: EMPATH proposes a multilingual auditor-judge benchmark for emotional-support chatbot safety (cs.AI updates on arXiv.org) · arxiv.org — LLM Evals Camilo Chacón Sartori
• 2026-07-01: Probability calibration reduces evaluator bias in LLM agent feedback loops (cs.CL updates on arXiv.org) · arxiv.org — LLM Evals Agents GLM5.2
• 2026-07-01: Probability calibration may reduce preference coupling in LLM evaluator feedback loops (cs.LG updates on arXiv.org) · arxiv.org — LLM Evals arXiv · GLM5.2
• 2026-07-03: ContextCodeCache generates a fresh codebase index for agents (Hacker News) · github.com — Codebase Indexing Context Engineering OpenAI · Anthropic · DeepSeek · Claude
• 2026-07-06: Price per 1M tokens is meaningless (Hacker News) · janilowski.pl — LLM Evals OpenAI · Anthropic · Artificial Analysis · GPT-4o · GPT-4 · GPT 5.5 · Claude Opus 4.8 · Sonnet 5 · GLM-5.2 · Fable 5
• 2026-07-10: A build-off compares GPT-5.6, Grok 4.5, Claude, Muse Spark, and open-weight models on four app tasks (Hacker News) · tryai.dev — Code Agents LLM Evals OpenAI · Anthropic · Meta · Fireworks · GPT-5.6 · Grok 4.5 · Claude Opus 4.8 · Claude Fable 5 · Muse Spark 1.1 · GPT 5.5 Qwen 3.7-Plus · kimi-k2.6 · GLM-5.2
• 2026-07-14: EvoClawBench tests whether agents can turn their own runs into reusable skills (cs.LG updates on arXiv.org) · arxiv.org — Agents LLM Evals Tool Use arXiv · GPT-5.4 · MiniMax-M2.7
• 2026-07-14: EYT-Bench benchmarks multi-turn dialogue with decoupled simulation, modeling, and judging (cs.CL updates on arXiv.org) · arxiv.org — LLM Evals Gemma 4 · GPT 5.5
• 2026-07-14: Agentic search sharply reduces unsupported links in construction-code QA (Все статьи подряд / Искусственный интеллект / Хабр) · habr.com — Agents Tool Use RAG LLM Evals RAG Evaluation Context Engineering DeepSeek · GPT 5.5 · Claude Opus 4.8 · Claude Sonnet 5 · Gemini-3.1-Pro
• 2026-07-16: STOCKTAKE: A benchmark for separating perception from action in LLM agents (cs.AI updates on arXiv.org) · arxiv.org — Agents LLM Evals Claude Sonnet 5 · GPT-5.4 · Grok 4.5
• 2026-07-16: Moonshot AI launches Kimi K3 and the limits of the pelican benchmark (Simon Willison’s Weblog) · simonwillison.net — LLM Evals Moonshot AI · Artificial Analysis Arena.ai · OpenRouter · DeepSeek · Anthropic · OpenAI · llama.cpp · LM Studio Simon Willison · Kimi K3 · kimi-k2.6 · Claude Opus 4.8 Max · Claude Fable 5 · GPT 5.5 high · GPT-5.6 Sol · GLM-5.2
• 2026-07-17: Shanghai AI Lab Introduces Intern-S2-Preview-397B for Scientific AI (量子位) · qbitai.com — Agents Tool Use LLM Evals 上海人工智能实验室 昇腾计算生态 · WAIC 2026 Intern-S2-Preview-397B Mobius · AlphaFold3 Rosetta Kimi-2.7-Code · GLM5.2 SWEBench-Pro SWEBench-Multilingual TerminalBench2.1
• 2026-07-17: US-China AI competition is splitting into open-model expansion versus controlled access (Все статьи подряд / Искусственный интеллект / Хабр) · habr.com — Open Source LLMs Long Context LLM Evals Google DeepMind · Reuters · Moonshot · Anthropic · White House World Artificial Intelligence Conference World AI Cooperation Organisation Pax Silica AI Opportunity Statement BRICS ASEAN United Nations Chinese state media Demis Hassabis Xi Jinping Trump · Kimi K3 Claude Fable 5 Max · GPT-5.6 Sol · Opus 4.8 · GPT 5.5 · R1
• 2026-07-20: Import AI 465: open-weight cyber gaps, Kimi K3, and Demis’ policy plan (Import AI) · importai.substack.com — LLM Evals Import AI UK government AI Security Institute AISI · DeepSeek · Kimi Demis · GLM-5.2 · Claude Opus 4.6 Opus 4.5 · GPT-5 · Claude Opus 4.5 · Sonnet 4.5 · Kimi K3
• 2026-07-22: GigaToken claims ~1000x faster language model tokenization (Hacker News) · github.com — Hugging Face · AMD · Apple Qwen/Qwen3-8B · LLaMA-3 · Llama 3.1 · Llama 3.2 · Llama 3.3 Qwen 2 · Qwen 2.5 · Qwen 3 · DeepSeek-V3 · DeepSeek-V3.1 · DeepSeek-V3.2 · DeepSeek R1 · DeepSeek-V4-Flash GLM-4 GLM 4.1V GLM-4.5 · GLM-4.7 · GLM-5 · GLM-5.2 · GLM-4.7-Flash Nemotron 3 · Nemotron 3 Nano · Nemotron 3 Super · Nemotron 3 Ultra · Kimi K2 · kimi-k2.5 · kimi-k2.6 · Kimi K2.7 Phi-4-mini Phi-4-multimodal · TinyLlama · Phi-3
• 2026-07-22: Testing Whether AI Labs Are Optimizing for the Pelican Benchmark (Hacker News) · dylancastillo.co — LLM Evals OpenRouter · GitHub · Hacker News Simon Willison · GPT-5.6 Terra · Claude Sonnet 5 · Gemini 3.5 Flash · Grok 4.5 · Qwen3.7-Max · GLM-5.2 · GPT-5.6 Luna Gemini 3.1 Flash-Lite · Claude Fable 5
• 2026-07-24: Opus 5 tops the Artificial Analysis Intelligence Leaderboard (Hacker News) · artificialanalysis.ai — LLM Evals Agents Tool Use Code Agents Long Context Artificial Analysis · Claude Opus 5 · Claude Fable 5 · GPT-5.6 Sol Mercury 2 HyperNova 60B 2605 Granite 4.0 H Small Gemma 3n E4B Instruct Nova Micro Sarvam 30B · Gemini 2.5 Flash Lite · Command A+ · Gemini 2.5 Flash · GLM-5.2 · MiniMax M3
• 2026-07-27: Zhijing: Engineering Social Intelligence as Testable and Trainable AI Capability (量子位) · qbitai.com — LLM Evals Agent Memory Chinese Academy of Sciences GPT 5.5 Zing-27B Zing-32B Zing-8B Zing-14B

Incident radar

GROUNDING’s Hallucination Incident Index has flagged 1 incident naming DeepSeek V4 Pro — a heuristic proxy over the AI-news radar’s daily journal (hallucination/jailbreak/refusal/bias keyword matches), not a verified incident registry.

• Severity: S2 1
• Category: Hallucination 1

Most recent:

• Agentic search sharply reduces unsupported links in construction-code QA source (2026-07-14)

Full breakdown: Hallucination Incident Index · Subscribe: incident-index RSS feed

FAQ

What is DeepSeek V4 Pro?

DeepSeek V4 Pro is the high-capability variant of DeepSeek’s V4 series previewed April 2026, a mixture-of-experts model with about 1.6T total and 49B active parameters and a 1-million-token context window. It is positioned near the frontier at low cost with open weights. GROUNDING tracks DeepSeek V4 Pro’s benchmarks, pricing, and availability.

What does this page track?

Dated radar mentions, source links, related concepts, and builder-relevant context for DeepSeek V4 Pro, collected automatically by GROUNDING.

When was DeepSeek V4 Pro last mentioned?

DeepSeek V4 Pro was most recently mentioned in a radar update dated 2026-07-27.

Category: Text / Language Models