DeepSeek V4 is widely expected to be DeepSeek's next flagship coding‑focused AI model, but as of early March 2026 most information about it comes from leaks, secondary reporting, and technical previews rather than an official launch announcement.
Reports point to a Lunar New Year–era release window, a new "Engram" memory architecture, and trillion‑parameter MoE scale aimed at long‑context coding and agentic workflows, but many benchmarks and exact specs are still best treated as rumors.
Quick overview: what is DeepSeek V4?
DeepSeek is a Chinese AI startup that gained global attention with its V3 and R1 models, which matched or beat Western systems at much lower training cost and briefly triggered a major tech‑stock sell‑off in early 2025. V4 is described in coverage and leaks as the fourth‑generation general model in this family, with a particular emphasis on software engineering, repo‑level reasoning, and long‑context tasks.
- A next‑gen architecture (often referred to as MODEL1/Engram) that reorganizes memory, cache, and attention for long contexts and cheaper inference.
- A strong focus on code: repo‑level understanding, multi‑file reasoning, and autonomous refactoring across large codebases.
- Open‑source release with permissive licensing, continuing DeepSeek's pattern with V3 and R1 and positioning V4 as a high‑end but self‑hostable coding model.
What seems relatively well‑supported (but still not "official")
Across multiple independent reports, a few technical themes keep showing up; while not "confirmed press‑release facts," they are better‑supported than one‑off forum rumors.
1. New MODEL1 / Engram architecture
Technical guides describe DeepSeek V4 as introducing four major innovations: MODEL1 tiered KV‑cache architecture, sparse FP8 decoding, Engram memory modules, and an mHC training optimization.
- Tiered KV cache (MODEL1): KV states are split across GPU VRAM, CPU RAM, and disk, reportedly reducing GPU memory use by about 40%, enabling context windows beyond 1 million tokens, and cutting serving costs significantly.
- Engram memory modules: Context (current conversation) is decoupled from long‑term "memory" stored in a vector database, allowing the model to retrieve relevant long‑term information without carrying the entire history in the prompt.
Multiple sources echo the idea that Engram allows "near‑infinite context retrieval" and persistent user‑ or project‑level memory, especially for coding workflows.
2. Scale and efficiency targets
Reports claim that DeepSeek V4 uses a 1‑trillion‑parameter Mixture‑of‑Experts design, with only a subset of experts active per token ("Top‑16 routed MoE"), targeting more than 80% on SWE‑bench Verified at 10–40× lower inference cost than Western competitors.
- Context window around 1 million tokens.
- Designed to run practical variants on dual RTX 4090/5090‑class hardware for on‑prem enterprise use.
3. Coding and repo‑level reasoning focus
Multiple sources frame V4 primarily as a coding model:
- Model previews describe V4 as "specialized for coding and software development," highlighting repo‑level reasoning where the model can understand and modify relationships across multiple files.
- Reports suggest internal benchmarks indicate V4 surpasses Claude and GPT series on long‑context code generation and multi‑file refactoring tasks.
What is clearly rumor or unverified
- Benchmark leaks: Unverified reports suggest V4 scores about 90% on HumanEval (above Claude and GPT‑4) and more than 80% on SWE‑bench Verified, but these numbers are not yet independently confirmed.
- Multimodal extensions: Claims that V4 will launch with image and video generation, but as of March 2 there is no widely available article or DeepSeek announcement specifying full multimodal support or exact timing.
- Exact release date: Several outlets repeat a mid‑February 2026 window around Lunar New Year, often attributing this to Reuters and The Information, but all also note that DeepSeek has declined to confirm a specific date and that one or more rumored windows have already passed.
- YouTube "leak analysis": Creator‑driven videos talk about V4 combining hardware‑level optimizations, architecture upgrades, and full multimodality in a "massive drop," but these are commentary and speculation, not primary sources.
Until DeepSeek publishes a formal announcement or open‑weights release, any specific benchmark score, parameter count, or launch day should be treated as rumor rather than fact.
Timeline tree of DeepSeek V4 news and rumors
Below is a timeline that shows how information about DeepSeek V4 has emerged, branching into confirmed reporting vs rumor/speculation at each step.
V3 (Dec 2024) and R1 (Jan 2025) release as open models with MoE architectures and very low reported training cost, competing with GPT‑4‑class models.
R1's success and low cost trigger a tech‑stock sell‑off in late January 2025, putting DeepSeek on the global AI map.
DeepSeek ships incremental updates like V3.1 and V3.2, expanding context windows but not yet introducing a clear next‑gen architecture.
Rumors start about an R2 model and future architectures focused on reasoning and long‑context, setting expectations that a major generational jump is being prepared.
Several trackers summarize a Reuters report that DeepSeek plans a new AI model focused on coding, targeted for February 2026; this is the first widely cited, named source tying V4 to a February launch window.
Main fact (coding‑focused model, February window) is widely re‑reported and treated as credible.
Precise date and configuration are not confirmed by DeepSeek and remain tentative.
Sites write that DeepSeek is preparing a new cutting‑edge model, probably timed to Chinese New Year, and note that prior major announcements like R1 also clustered around that period.
Confirms industry expectation of a Lunar New Year release pattern.
Still no official name "V4" in DeepSeek's own communication; attribution comes from external analysis.
DeepSeek publishes a paper signed by founder Liang Wenfeng introducing "Conditional Memory" and the Engram retrieval architecture, decoupling long‑term memory from immediate context.
On the same date, blogs start explicitly labelling this as the core of "DeepSeek V4," predicting that the production model will use Engram for 1M+ token contexts and repo‑scale code understanding.
The Engram paper itself is real and clearly from DeepSeek; conditional memory is not rumor.
Mapping Engram directly to a product called "V4" and its exact benchmarks remains extrapolation.
Detailed guides describe MODEL1 tiered KV cache (40% memory reduction, 10× context extension), sparse FP8 decoding (1.8× speedup), and Engram long‑term memory as if these are established attributes of V4.
Blogs synthesize industry expectations: early‑2026 release window, focus on reasoning stability, long‑context reliability, and production‑grade engineering workflows.
The architectural ideas (tiered KV cache, sparse FP8, Engram, mHC) are backed by technical write‑ups, though sometimes using "V4" as a label before DeepSeek brands it that way.
Claims about exact cost reductions, performance multipliers, and benchmark leads are still based on DeepSeek or partner internal numbers rather than public evaluation.
Multiple outlets state that DeepSeek is targeting mid‑February 2026, often around February 17, to launch V4, mirroring R1's timing one week before Lunar New Year in 2025.
Reports note DeepSeek "declined to comment on specific release timing," despite widespread assumptions of a mid‑February debut.
It is widely reported that industry insiders expect a February launch and that DeepSeek is working on a coding‑centric successor.
The named dates pass without a public open‑weights V4 release or major launch event as of late February.
Outlets publish explainers summarizing: V4 expected around Lunar New Year, four major technical innovations, and a strategic shift toward a China‑focused Cursor‑like coding ecosystem.
These pieces also note that DeepSeek's open‑source market share dropped from about 50% to under 25% in 2025 amid competition from Qwen, Kimi K2, and others, framing V4 as an attempt to regain leadership.
Release trackers log that a second rumored launch window passes without release, and that DeepSeek remains silent on the delay.
The same tracker adds unverified benchmark leaks (HumanEval ~90%, SWE‑bench Verified >80%), explicitly labelling them as "internal claims pending independent verification."
Node 1: "V4 is real and coding‑focused" – high confidence from multiple sources.
Node 2: "Exact scores and cost multipliers" – still rumor, driven by internal or partner benchmarks.
Blog posts categorize "rumors vs reality" around V4, focusing on Engram + mHC as the likely core and speculating about local privacy via dual‑GPU on‑prem deployments.
Reddit threads claim that reports suggest V4 will launch "next week" with image and video generation, but these threads do not link to a widely accessible article or DeepSeek confirmation.
YouTube creators release "DeepSeek V4 update" videos talking about DeepGEMM updates, Blackwell GPU support, and a "massive new drop," but again these are commentary layered on top of existing reports.
Aggregators still list DeepSeek V4 as "upcoming," with an early‑2026 release window and no confirmed public release date.
Technical pieces continue to describe MODEL1/Engram, sparse FP8, and trillion‑parameter MoE design as "V4's architecture," but the model itself has not yet been independently benchmarked in open competitions or widely distributed.
Why DeepSeek V4 matters if the leaks are right
If the more credible reports hold up, DeepSeek V4 could matter in several ways:
- Economics of agents and long‑context coding: Tiered KV cache and sparse FP8 decoding aim to make million‑token contexts and repo‑scale reasoning economically viable for production agents, not just demos.
- Open‑source pressure on proprietary models: An Apache‑ or MIT‑licensed trillion‑parameter MoE coding model with top‑tier benchmarks would intensify competitive pressure on closed systems like GPT and Claude, especially for enterprises that require on‑prem deployment.
- Strategic shift to tools: DeepSeek is positioning V4 as the backbone of a "model + tools" ecosystem (e.g., a Cursor‑like coding environment for Chinese developers), echoing a broader industry move from raw models to full platforms.
How to read the news (and rumors) going forward
When you see new DeepSeek V4 headlines over the next few weeks, it helps to bucket them:
- • Papers or tech reports directly linked to DeepSeek (e.g., Engram/conditional memory descriptions).
- • Major outlets citing named sources like Reuters, The Information, or direct interviews when discussing release windows and strategic goals.
- • Single‑source benchmark screenshots, especially without evaluation details.
- • Exact launch dates, or claims of sudden multimodal capabilities, that trace back only to forums or unlinked paywalled articles.
- • YouTube or social posts that talk about "massive drops" without adding new primary evidence.