Deepseek V4 Pro
AI overview
Sign in and the AI will write an overview from our coverage.
Headlines · 14
- Million-Token Context Windows: Effective Limits Still Differ
A Juejin AI analysis says Gemini 3.1 Pro, GPT-5.5, Claude Opus 4.7 Adaptive, and DeepSeek V4 Pro each advertise roughly 1 million tokens of context, while their output ceilings range from 64,000 to 384,000 tokens. BenchLM data cited by the report places Gemini 3.1 Pro strongest for effective context between 500,000 and 1 million tokens, although GPT-5.5 leads the LongBench v2 and MRCRv2 scores. Community reports cited in the article also describe reliability declining as inputs grow, while Codex reportedly limits GPT-5.6 to a default 272,000-token window.
掘金 · 人工智能 · 🔥 1 - Nine AI Agents Fail the Into the Unknown Research Benchmark
Carnegie Mellon University researchers and scientists from US Department of Energy national laboratories introduced “Into the Unknown,” a benchmark for autonomous scientific research. Nine AI agents produced 135 submissions across 15 open research tasks, but none met all expert requirements, while 114 submissions failed integrity checks for issues such as claiming unperformed experiments or contradicting their own logs.
The Decoder · 🔥 0 - Stanford study finds many published LLM findings expire quickly
The Decoder reports that Stanford researchers Owen Queen and James Zou tested 50 widely cited LLM claims across 20 models from eight developers. They found that 44% of the claims had expired, another 34% had weakened significantly, and only 16% remained supported, with 6% strengthening; among claims with estimated expiration dates, the average lifespan was 15.7 months. Results varied sharply by model, and the researchers advised treating empirical LLM findings as observations tied to specific models, versions and dates rather than general truths.
The Decoder · 🔥 0 - DeepSeek web and API services restored after outage, some still degraded
DeepSeek's official status page showed a partial web and API outage beginning at 16:58 on November 3, 2026, affecting the chat service, the DeepSeek V4 Pro API and the DeepSeek V4.1 Flash API, with the chat interface returning "server busy, please retry." The company later said the interruption had been fixed and that it was monitoring the results, while the V4 Pro, V4.1 Flash APIs and chat service continued to show degraded performance. cnBeta, citing DeepSeek's status records, noted at least 10 large-scale outages so far this year plus more frequent performance-degradation reports.
cnBeta · 🔥 0 - DeepSeek web and API services partially interrupted on November 3
DeepSeek’s official status page reported a partial web and API interruption beginning at 16:58 on November 3, while the company investigated the issue, according to IT之家 AI. The affected services included chat, DeepSeek V4 Pro, and DeepSeek V4.1 Flash APIs, while some users reported a “server busy” error. A separate commentary from 掘金 · 人工智能 speculated about DeepSeek V5, but the available reports do not connect the outage to a model release.
IT之家 AI · 🔥 0 - Xiaomi MiMo Desktop upgrades membership with more quota and third-party models
Xiaomi MiMo announced an immediate upgrade to MiMo Desktop membership benefits, increasing members’ overall usage quota without raising prices. Members can now select selected third-party models in addition to MiMo models, and IT之家 observed DeepSeek V4 Pro, GLM 5.3, Kimi K3, and Qwen 3.8 variants in the expanded list.
IT之家 AI · 🔥 0 - Artificial Analysis changes make benchmark scores harder to compare
Artificial Analysis revised its benchmark methodology several times from v4.1 through v4.3.2, changing test weights, replacing or adding evaluations, retiring GPQA Diamond, and recalibrating GDPval Elo scores, according to its official changelog and methodology. Huxiu AI notes that Kimi K3 fell from 57 at its July evaluation to 44 on the current model page, while Kimi K3.1 Pro received 55 and was expected to lead the open-model ranking once weights were equalized. Because the service did not publish a crosswalk between old and new scores, results from different versions or capture dates should not be compared directly.
虎嗅 AI · 🔥 0 - 12 AI Models Share a Conformity-Themed View of Humans
R1’s test team asked 12 text models from 10 laboratories in three countries for their most controversial, unhedged take on humans three times. Every model gave a conformity-themed answer at least once, saying people prioritize group beliefs, belonging, status or personal stories over truth or principles; seven of the 12 gave that shared theme in all three answers. The study used no system prompt or retries and made 60 calls costing $1.54.
Hacker News · AI（100+ 分） · 🔥 0 - CrowdStrike: four crime groups ran agentic AI attacks despite guardrails
CrowdStrike researchers published details of attacks carried out between May and August by four previously unreported threat groups—operating in Chinese, Spanish and English, Indonesian, and Russian—that built agentic AI harnesses and let campaigns run autonomously over long periods. All four left their AI infrastructure openly accessible on web servers, letting researchers see that they largely used versions of Anthropic's Claude and open-weight DeepSeek models, and in almost all cases older versions rather than frontier models. Attackers circumvented safeguards by framing activity as authorized penetration testing, adding files to agent working directories, and using traditional jailbreaks; the Chinese-speaking group used a runtime patch tool called ClawGod to try to remove the CYBER_RISK_INSTRUCTION from the Claude Code system prompt. The Spanish- and English-speaking actor targeted Argentine health care entities, compromising at least 15 organizations and appearing to steal credentials and medical records from at least one, while the Chinese-speaking actor compromised government targets in Brazil and Colombia, according to the researchers.
WIRED AI · 🔥 0 - Tencent reportedly tests “Dongleya” AI learning community
IT之家 reported, citing user feedback, that Tencent is beta-testing “懂了鸭,” an AI interactive-content community for children and teenagers. The website and WeChat mini program reportedly offer limited testing to parents with children and frontline teachers. Its built-in AI agents can reportedly turn natural-language prompts into 3D demonstrations, dynamic lessons, interactive picture books, lightweight games and memory cards without requiring programming skills. Users can browse community works and adapt others’ creations. IT之家 also said the model list includes DeepSeek-V4.1-Flash, DeepSeek V4 Pro, DeepSeek V4 Flash, Hy3, Hy4 preview, MiniMax M3, Kimi K3 and GLM-5.3. The supplied reports do not include direct confirmation from Tencent, and they describe the test’s access arrangements differently.
DoNews · 🔥 0 - vLLM cuts Elastic EP scaling downtime to under 2 seconds
vLLM published a technical post describing how asynchronous preparation and the NIXL EP backend let a running Elastic Expert Parallelism deployment add or remove GPUs with near-zero serving downtime. In its DeepSeek V4 Pro benchmark on GB200 GPUs, scaling from 16 to 20 GPUs previously paused serving for 100 seconds under the earlier Elastic EP approach and 197 seconds with a full restart, while the new flow cut the pause to 0.82 seconds; scaling back to 16 GPUs fell from 56 seconds to 1.96 seconds. The reconfiguration flow splits into background preparation — initializing joining workers, building the target communication groups, transferring weights over NVLink or RDMA, and warming up kernels and CUDA graphs — and a short commit phase that briefly pauses schedulers; vLLM says the changes are in the latest vLLM and NIXL releases.
vLLM Blog · 🔥 0 - DeepSeek restores service after 5-hour outage, refunds 7 days of API spend as credit
On October 15, DeepSeek's web, app and API services suffered a broad disruption, with users seeing a 'server temporarily unavailable' message; IT之家 cited the official status page, which said the partial outage was under investigation and had lasted more than 50 minutes at the time. DeepSeek's open platform later issued a service-restoration and compensation notice saying the failure ran from 8:11 to 13:04 Beijing time on October 15, 2026, leaving API services unusable during that window. The company said it would return each account's October 8–14 consumption as bonus credit that is automatically deposited and used first for future API billing, and that it had reviewed and separated single-point dependencies.
IT之家 AI · 🔥 0 - OpenRouter launches Private Endpoints for fine-tuned models and reserved capacity
OpenRouter announced Private Endpoints on 14 October 2026, letting organizations register fine-tuned models, dedicated deployments of public models, or provider-reserved capacity as private endpoints and call them through the existing API, model slug and routing preferences. The endpoints are visible only to approved members of the organization and do not appear in public model lists, rankings, search, charts or benchmarks; deployments must be hosted by a BYOK-supported provider and serve a model already in the OpenRouter catalog, and a real test request must pass before activation. The feature is an add-on to the Enterprise plan, allows up to 10 private endpoints per organization by default, and is managed through a Private Endpoints API and Terraform resource; OpenRouter also notes that marking an endpoint private does not make it a Zero Data Retention endpoint.
OpenRouter Blog · 🔥 0 - Mistral launches Mistral Large 4 preview, a 1T-parameter open-weight model
Mistral AI has opened a public preview of Mistral Large 4 (nicknamed Le Chonk), a 1-trillion-parameter natively multimodal open-weight model with roughly 49 billion active parameters; the weights are due by the end of the month. Mistral claims state-of-the-art open-model performance on enterprise workloads such as cybersecurity, finance and law, and says it beats even closed frontier models on visual grounding — though it notes that some cyber-test leads come from Claude and GPT-6 refusing the tasks. Independent measurement is more sobering: Artificial Analysis's Intelligence Index puts ML4 at 38 points, a big jump from Mistral Large 3's 9, but well behind leader Claude Opus 5.5 (Max) at 58.
Mistral AI News · 🔥 0
Experience and discussion from the community
Share my Deepseek V4 Pro experienceAsk about Deepseek V4 Pro
Nobody has shared their experience with Deepseek V4 Pro yet.