DeepSeek V4 ProおよびFlashモデル、Huawei Ascendチップで動作可能に

After a couple months’ delay and lots of speculation, DeepSeek finally released the heavily anticipated DSV4, the first major version model since DSV3 (Dec 2024) and DSR1 (Jan 2025). It brings the DeepSeek family up in line with Kimi K2.6, the current open model leader, and Xiaomi Mimo 2.5, a lesser known family released 2 days ago.

The DSV4 family is roughly a Gemini 3.1, GPT 5.4, Opus 4.6 level model, up to 1.6T MOE withtrained on 32T tokens with FP4, with 1M token context (supported by their new Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA) techniques), and incredibly rarely, they released both the Base and Instruct versions - surely setting the stage for a possible “DeepSeek R2” in future, though this one already has reasoning effort.

The technical report is a typically dense 58 pages, demonstrating training and inference insights and improvements from the Manifold Constrained Hyper-Connections (mHC) paper they released in January, continued usage of Moonshot’s Muon, and CSA/HCA’s overall INCREDIBLE efficiency improvements on DeepSeek 3.2-Exp’s already impressive Sparse Attention - at 1M tokens, requiring only 27% of FLOPs and 10% of KV cache memory compared with DeepSeek-V3.2:

The geopolitical backdrop behind the Huawei CANN compatibility is DeepSeek weaning dependence off export-controlled NVIDIA/CUDA chips — Ascends are still a quarter the supply of H100s, but this is an important milestone for Chinese total independence.

AI News for 4/23/2026-4/24/2026. We checked 12 subreddits, 544 Twitters and no further Discords. AINews’ website lets you search all past issues. As a reminder, AINews is now a section of Latent Space. You can opt in/out of email frequencies!

AI Twitter Recap

Top Story: DeepSeek V4

DeepSeek released DeepSeek-V4 Pro and DeepSeek-V4 Flash, its first major architecture refresh since V3 and first clear two-tier lineup, with 1M-token context, hybrid reasoning/non-reasoning modes, an MIT license, and a technical report detailed enough that multiple researchers called it one of the most important or best-written model papers of the year. Across the reactions, the factual consensus is that V4 materially advances open-weight long-context and agentic coding performance while remaining somewhat behind the top closed frontier models overall. Independent benchmarkers place V4 Pro around the #2 open-weights tier, roughly near Kimi K2.6 / GLM-5.1 / strong Claude Sonnet-class to Opus-ish depending on benchmark and mode, with especially strong long-context and agentic performance; opinions diverge on how close it is to GPT-5.x / Opus 4.7 and on whether this is “democratizing” progress or an architecture so complex that few open labs can realistically reproduce it. Key sources include deep-dive commentary from @ArtificialAnlys, @scaling01, @nrehiew_, @ben_burtenshaw, @TheZachMueller, @ZhihuFrontier, and infra/vendor posts from @vllm_project, @NVIDIAAI, and @Togethercompute.

Core facts and technical details

The most concrete technical claims repeated across the discussion:

Two models

V4 Pro: 1.6T total parameters / 49B active

V4 Flash: 284B total / 13B active

Reported by @ArtificialAnlys, @teortaxesTex, @baseten, @NVIDIAAI

Context

1M tokens, up from 128K in V3.2 per @ArtificialAnlys

Multiple posters frame this as the headline achievement: “solid ultra-long context” @teortaxesTex

Training scale

32T–33T tokens cited repeatedly

@nrehiew_ notes 32T tokens over 1.6T parameters, i.e. roughly 20 tokens/parameter

@teortaxesTex cites 33T

@nrehiew_ estimates pretraining compute at ~1e25 FLOPs

Reasoning / modes

DeepSeek exposes three reasoning modes per @Togethercompute

Hybrid “thinking/non-thinking” positioning noted by @ArtificialAnlys

Long-context architecture

Several threads summarize a new hybrid attention system:

shared KV vectors

compressed KV streams

sparse attention over compressed tokens

local/sliding-window attention for nearby context

@ZhihuFrontier gives the most compact public summary:

2× KV reduction via shared key-value vectors

c4a ≈ 4× compression

c128a ≈ 128× compression

top-k sparse attention on compressed tokens

128-token sliding window

1M context KV cache = 9.62 GiB/sequence (bf16)

8.7× smaller than DeepSeek V3.2’s 83.9 GiB

FP4 index cache + FP8 attention cache gives another ~2× reduction

@ben_burtenshaw condenses this to “10× smaller KV cache”

@TheZachMueller and @TheZachMueller describe CSA + HCA layer patterns, with alternating layers and V4 Flash using sliding-window layers instead of HCA in some places

Quantization / checkpoint format

@LambdaAPI: checkpoint is mixed FP4 + FP8

MoE expert weights in FP4

attention / norm / router in FP8

claim: the full model fits on a single 8×B200 node

Inference hardware / serving

@NVIDIAAI: on Blackwell Ultra, V4 Pro can deliver 150+ TPS/user interactivity for agentic workflows

@NVIDIAAI: published day-0 V4 Pro performance pareto using vLLM

@SemiAnalysis_: day-0 support and benchmarking across H200, MI355, B200, B300, GB200/300

@Prince_Canuma: DeepSeek4-Flash on 256GB Mac

@Prince_Canuma: MLX quants published

@simonw asks about smaller-RAM Mac viability, implying community interest but incomplete support story

@QuixiAI reminds users that many local stacks still lack tensor parallel, relevant because V4-class models strongly stress inference infra

License / availability / pricing

MIT license per @ArtificialAnlys

first-party API plus rapid third-party availability via @Togethercompute, @baseten, @NousResearch, @Teknium

V4 Pro pricing: $1.74 / $3.48 per 1M input/output tokens

V4 Flash pricing: $0.14 / $0.28

cache-hit pricing also given by @ArtificialAnlys

@scaling01 views the pricing as a glimpse of future “Mythos-level” cheap coding models

Reuters-via-posted quote from @scaling01: DeepSeek said Pro pricing could fall sharply once Huawei Ascend 950 supernodes are deployed at scale in H2

Independent evaluations and where V4 lands

The most useful independent benchmark synthesis came from @ArtificialAnlys:

V4 Pro Max: 52 on Artificial Analysis Intelligence Index

up 10 points from V3.2 at 42

becomes #2 open weights reasoning model, behind Kimi K2.6 (54)

V4 Flash Max: 47

positioned around strong mid/high open models, “Claude Sonnet 4.6 max level intelligence”

GDPval-AA (agentic real-world work):

V4 Pro: 1554, leading open-weight models

ahead of Kimi K2.6 (1484), GLM-5.1 (1535), MiniMax-M2.7 (1514)

AA-Omniscience

V4 Pro: -10, an 11-point improvement over V3.2

but still paired with 94% hallucination rate

V4 Flash: 96% hallucination rate

Cost to run AA Index

V4 Pro: $1,071

V4 Flash: $113

Output tokens used on AA Index

V4 Pro: 190M

V4 Flash: 240M

This is a major caveat: cheap per-token pricing does not imply cheap total task cost if the model spills huge token volumes

Additional eval perspectives:

@arena:

#2 open in Text Arena overall at debut

category wins/placements:

#1 Medical & Healthcare

#15 Creative Writing

#18 Multi-Turn

thinking variant:

#8 Math

#9 Life/Physical/Social Science

@arena emphasizes the Pro vs Flash tradeoff:

Pro ranks ~30 places higher

costs 12× more

Flash is still competitive in Chinese, medicine, math

@scaling01:

“~Opus 4.5 estimate holds for now, at least on SimpleBench”

@scaling01:

V4 is “definitely better than GLM-5.1 but not quite Opus 4.7, GPT-5.4 or Gemini 3.1 Pro”

@scaling01 lists what scores would confirm <6 month gap:

ARC-AGI-1 ~75%

ARC-AGI-2 ~35%

GSO ~26%

METR 4.5–5 hours

WeirdML ~63%

@TheZachMueller:

on his evals, Flash@max ≈ Pro@high on reasoning

Pro focuses more on knowledge (SimpleQA)

@VictorTaelin:

after fixing benchmark bugs and letting long-running models run longer, DeepSeek and Kimi improved materially

@mbusigin:

a simple negative early impression with no detail

@petergostev:

on BullshitBench, not about capability but refusal/pushback behavior, GPT-5.5 underperformed; included here because many readers compare V4 in an eval-skeptical environment

Facts vs opinions

Facts / relatively well-supported claims

V4 Pro / Flash were released with the specs above, MIT-licensed, 1M context, and open technical documentation: @ArtificialAnlys, @TheZachMueller

The architecture introduces a new long-context attention system with dramatic KV-cache reduction: @ZhihuFrontier, @ben_burtenshaw

Independent benchmarkers broadly place V4 Pro near the very top of open weights but below the best proprietary models overall: @ArtificialAnlys, @arena, @scaling01

DeepSeek V4 is heavily token-intensive in some evaluations: @ArtificialAnlys

The checkpoint uses FP4/FP8 mixed precision and can fit on an 8×B200 node: @LambdaAPI

Rapid ecosystem support arrived via vLLM and other providers day 0: @vllm_project, @SemiAnalysis_

Opinions / interpretation

“V4 is ~4–5 months behind the frontier” from @scaling01, @scaling01, @scaling01 is an informed estimate, not a measured fact

“Top three open” vs “only open model close to frontier” debate from @teortaxesTex is partly about benchmark trust and framing

“Strongest pretrained model we have” from @teortaxesTex is an opinion hinging on scale + architecture, not direct benchmark supremacy

“Most significant AI paper of the year” from @Dorialexander is enthusiasm, not consensus

“This is what research should look like” from @scaling01 speaks to transparency/style rather than only capability

“Not exactly a democratizing technology” from @teortaxesTex is a strong architectural/political interpretation

Different opinions and fault lines

1) Is V4 near frontier, or clearly behind?

More favorable

@scaling01: puts it at roughly GPT-5.2 / Opus 4.5+ tier

@scaling01: SimpleBench supports ~Opus 4.5

@teortaxesTex: argues it is the strongest pretraining base among opens and implies people are underestimating what post-training can do

More skeptical

@scaling01: below Opus 4.7 / GPT-5.4 / Gemini 3.1 Pro

@scaling01: the gap may widen again because closed labs have bigger models, better science/law/medicine coverage, faster inference with GB200s

@mbusigin: early impressions “not great”

@teortaxesTex: says polished models like K2.6 and GLM 5.1 may still feel better in coding despite lower intrinsic capacity

2) Is V4’s real contribution model quality, or long-context systems design?

A big split in reactions is that many technical readers think the long-context architecture matters more than the raw benchmark position.

@teortaxesTex: “They’ve completed their quest: Solid Ultra-Long Context”

@ben_burtenshaw: first open model where long context and agentic post-training “meet”

@scaling01: expects other open labs to adopt pieces of the architecture

@Dorialexander: frames Huawei/sovereignty constraints as an opportunity to reshape hardware and memory/interconnect design

@jukan05: reads the paper as evidence that NVIDIA’s hardware roadmap is unusually well aligned to where MoE/long-context models are going

3) Is V4 “open democratization,” or too hard to copy?

This was one of the sharpest strategic disagreements.

@teortaxesTex: says V4 is “not exactly a democratizing technology” because the architecture is too difficult for most labs to replicate

@teortaxesTex: suggests even DeepSeek may not want to do this exact architecture again without refactoring

@stochasticchasm: notes the sheer hyperparameter complexity is daunting

Against that, @Prince_Canuma and @Prince_Canuma show that the ecosystem is already compressing and adapting Flash for localish Apple Silicon use, softening the “not democratizing” claim on the inference side if not the training side

4) Are people underrating Flash?

Several reactions suggest Flash may be more important than Pro for practical adoption.

@arena: Flash shifts the price/performance frontier

@TheZachMueller: Flash@max ≈ Pro@high on reasoning tasks

@teortaxesTex: benchmarks may underweight “legit 1M context for pennies”

@Prince_Canuma: Flash runs on 256GB Mac

@baseten and @Togethercompute emphasize long-document analysis and agentic use cases where Flash’s economics matter

China, chips, Huawei, and sovereignty context

DeepSeek V4 was not discussed as a pure model release; it was treated as evidence in the larger US–China compute and sovereignty debate.

@scaling01: Chinese labs are already in or near “takeoff” in the sense that their models help build better models, though still shifted 5+ months behind

@scaling01: thinks chip bans are likely to widen the gap in broad domains over time

@teortaxesTex, @teortaxesTex: disputes simplistic Huawei-dismissal and notes mixed Chinese sentiment toward Huawei

@ogawa_tter: points to analysis of Ascend 950 / A3 clusters and V4 deployment plans

@Dorialexander: argues the sovereignty play around Huawei may reshape hardware architecture

@scaling01: cites DeepSeek saying prices could drop sharply once Ascend 950 supernodes scale in H2

@jukan05: interprets V4 as validating NVIDIA’s Blackwell/Rubin/HBM/interconnect strategy

@NVIDIAAI, @NVIDIAAI: unsurprisingly highlight Blackwell day-0 performance, but this is vendor framing rather than independent proof of strategic superiority

There is also a more ideological thread:

@teortaxesTex, @teortaxesTex, @teortaxesTex argues that Western discourse often misreads Chinese labs as purely state proxies or distillation shops, and instead sees them as serious mission-driven actors. This is interpretive, but it helps explain why the release drew such emotionally charged geopolitical reactions.

Distillation, training data, and data quality

A recurring undercurrent: does V4 mainly reflect architectural innovation, or can critics dismiss it as “distillation”?

@yacineMTB speculates that some complaints about Chinese distillation may partly come from people discovering they’re outperformed

@cloneofsimo: “Very interesting... given they distilled claude 樂樂”

@kalomaze: jokes about DeepSeek training on DeepSeek reasoning traces

On the more substantive side, @teortaxesTex says DeepSeek’s writing quality, especially Chinese, reflects long-standing obsession with data cleanliness and cites job listings @teortaxesTex, @teortaxesTex

@nrehiew_ notes the report still lacks much detail on pretraining data beyond standard categories

Overall, factual public evidence in this tweet set supports “DeepSeek trains at large scale with strong data work,” but not any strong claim about the degree of external distillation beyond speculation

Architecture lineage and prior art

Several researchers pointed out that V4 did not emerge from nowhere.

@jaseweston: says DeepSeek uses hash routing from a 2021 ParlAI approach

@suchenzang: criticizes routing-induced outliers, with a jab at hashing

@teortaxesTex: notes Mixtral-style MoE was a reasonable earlier hack, but claims DSMoE changed things

@art_zucker broadly attacks MoEs as a dead end

@gabriberton counters that MoEs are provably effective despite inelegance

@stochasticchasm is even more positive: “MoEs are amazing”

This matters because V4 was read not just as a stronger checkpoint, but as a possible new design point for open long-context MoEs.

Why the technical report itself mattered

A striking amount of praise was directed not just at the model but at the paper/report quality.

@scaling01: “the technical paper is a big deal”

@Dorialexander: “most significant AI paper of the year”

@morqon: “one of the best I’ve ever read”

@scaling01: “this is what research should look like”

@TheZachMueller, @iamgrigorev, @nrehiew_: all signal unusually high effort to digest and test the report

For expert readers, this is important because many frontier releases now arrive with sparse technical disclosure. V4’s report appears to have reset expectations for what a serious open release can look like.

Practical limitations and caveats

Despite the enthusiasm, several caveats recur:

Still behind closed frontier in aggregate capability

especially sciences/law/medicine and broad “general domains” per @scaling01

Reasoning RL may be undercooked

@scaling01: reasoning efficiency not much changed vs V3.2 Speciale

Serving remains hard

@scaling01: many labs serve at only 20–30 tok/s and limited concurrency; running evals can take a day

@ClementDelangue: acknowledges concurrency bottlenecks on HF

High token usage

major practical caveat from @ArtificialAnlys

API controls

@stochasticchasm: notes DeepSeek API appears not to allow sampler control

Adoptability

@teortaxesTex: too complex for many labs to copy cleanly

Broader implications

Three implications stand out.

Open-weight long-context is no longer just marketing.

V4’s strongest contribution may be proving that 1M context can be made operationally credible in an open-weight model, with concrete KV-cache engineering and open inference support. This is why multiple posters focused less on benchmark deltas and more on systems design: @ben_burtenshaw, @ZhihuFrontier, @scaling01.

China’s top labs remain competitive in open models, even if not fully closing the closed-model gap.

The benchmark picture across @ArtificialAnlys, @arena, and @scaling01 suggests Chinese labs now dominate much of the open-weight top tier: Kimi, GLM, DeepSeek, and soon MiMo.

The bar for “open” is rising from checkpoint release to full-stack co-design.

V4 was instantly discussed alongside vLLM, Blackwell, MLX quants, Mac viability, Ascend clusters, and cache/memory architectures. In other words, “the model” is increasingly inseparable from the inference substrate.

Infrastructure, inference, and local/open ecosystem

Hugging Face launched ML Intern, an open-source CLI “AI intern” for ML work that can research papers, write code, run experiments, use HF datasets/jobs, search GitHub, and iterate up to 300 steps, per @MillieMarconnni. Related sentiment: HF’s $9 Pro tier is unusually strong value per @getpy.

Meta said it will add tens of millions of AWS Graviton cores to its compute portfolio to scale Meta AI and agentic systems for billions of users, per @AIatMeta.

Local/open coding stack momentum stayed strong:

@julien_c: Qwen3.6-27B via llama.cpp on a MacBook Pro feels close to latest Opus for many coding tasks

@p0: free CLI agent built with Pi + Ollama + Gemma 4 + Parallel web search MCP

@Prince_Canuma: DeepSeek V4 quants incoming

@QuixiAI: reminder that llama.cpp / Ollama / LM Studio do not support tensor parallel, pushing serious multi-GPU serving users toward vLLM

Nous/Hermes shipped heavily:

Hermes Agent v0.11.0 introduced a rewritten React TUI, dashboard plugin, theming, more inference providers, image backends, and QQBot support, per @WesRoth

Hermes got broad praise and rapid support for both DeepSeek V4 and GPT-5.5, via @mr_r0b0t, @Teknium

@JulianGoldieSEO and @LoicBerthelot compared Hermes favorably to OpenClaw on learning loops, memory, model support, deployment flexibility, and security

A native Linux sandbox backend for Deep Agents using bubblewrap + cgroups v2 was released by @nu_b_kh

Research papers and benchmarks

On-policy distillation token selection:

@TheTuringPost highlights a paper showing only some tokens carry most learning signal; using ~50% of tokens can match or beat full training and cut memory by ~47%, while even <10% focused on confident-wrong tokens nearly matches full training.

Google Res