DeepSeek V4 lands with 1M context

// 45d agoMODEL RELEASE

DeepSeek V4 lands with 1M context

DeepSeek has previewed V4 Pro and V4 Flash, a new MoE model family with 1M-token context windows and stronger coding, reasoning, and agentic performance claims. The release keeps DeepSeek in the frontier-model race, but the most meaningful numbers still come from DeepSeek’s own benchmarks and need outside validation.

// ANALYSIS

This is a real step up in packaging and efficiency, not just another benchmark tweet.

–V4-Pro and V4-Flash both support 1M-token context, with Pro at 1.6T parameters and 49B activated, and Flash at 284B/13B.
–DeepSeek says the architecture cuts long-context compute and KV-cache costs versus V3.2, which is the part that matters for practical deployment.
–The API docs already map `deepseek-chat` and `deepseek-reasoner` to V4-Flash behavior, so this is already entering the product surface.
–The release is still a preview, so independent evals matter more than the company’s top-line benchmark claims.

// TAGS

deepseek-v4llmreasoningagentopen-sourceapi

DISCOVERED

45d ago

2026-04-24

PUBLISHED

45d ago

2026-04-24

RELEVANCE

10/ 10

AUTHOR

Alternative-Duty-532

// KEEP READING

More AI developer news from the feed

EXPLORE FULL FEED

BENCHMARK7m ago

Structured docs boost coding agent precision

Mintlify published results showing that connecting Claude Code to structured documentation via MCP on a million-line monorepo improved answer precision by 64%. The experiment also demonstrated a 50% reduction in token consumption and a 1.5x speedup in task completion.

VIDEO47m ago

Mastra outlines transition to structured specifications

In this episode of the AI Agents Hour, Mastra co-founders Shane Thomas and Abhi Aiyer explore shifting from manual prompting to structured, developer-first agent specifications. The discussion covers tool integrations like CodeRabbit for pull request reviews and Freestyle's forkable Linux VM sandboxes to build reliable, deterministic AI agents.

BENCHMARK1h ago

Cognition has introduced FrontierCode, a software engineering benchmark that measures the mergeability, quality, and maintainability of AI-generated code rather than simple functional correctness.

Cognition has launched FrontierCode, a new software engineering benchmark designed to evaluate whether AI-generated code meets the high standards of production-grade codebases. Developed in partnership with the maintainers of 36 flagship open-source repositories who spent over 40 hours per task, the benchmark consists of three nested subsets (Extended, Main, and Diamond) totaling 150 tasks. FrontierCode goes beyond classical unit tests to measure mergeability, style, regression safety, and scope using novel grading methods like reverse-classical testing (which ensures agent-written tests fail on the buggy base code) and adaptive classical grading with their new tool, mutagent. In initial testing, Claude Opus 4.8 led all models but only scored 13.4% on the hardest Diamond set, highlighting that frontier models still struggle significantly with writing production-ready, maintainable code.