deepseek-v4-pro
AI overview
Sign in and the AI will write an overview from our coverage.
Headlines · 11
- WEKA and NAND Research launch Inference Economics Scorecard
WEKA and NAND Research launched the Inference Economics Scorecard at Supercomputing 2026, with nine metrics for assessing inference productivity and economic value beyond GPU utilization. The framework requires at least eight-hour test runs, counts only successful outputs, and was used in an eight-hour WEKA-Nebius benchmark on an NVIDIA HGX B300 system. WEKA reported that its NeuralMesh configuration produced 3.7 times more output tokens per GPU-hour than the HBM-only setup, though the performance results have not been independently verified here.
科技新报 · 🔥 3 - MatOS opens source for AI-assisted materials research
Beijing Zhongguancun Academy and the Zhongguancun Institute of Artificial Intelligence have open-sourced MatOS, an agent system for materials research. Its asset library contains 3,135 datasets, 386 skills, 104 workflows and 51 expert roles, while the system integrates 48 materials-computing applications. The project team reports that MatOS led compared systems on ResearchClawBench and MatTools, but the supplied reports provide no independent replication.
36氪 人工智能 · 🔥 0 - Qianwen Office Adds 10 Models and Opens Model Selection
PingWest and ZhidxcomAI report that Qianwen Office launched 10 models, including Qwen4-Max and third-party models from Kimi, GLM, DeepSeek, MiniMax and Step. Personal and enterprise-standard users can choose models according to task complexity, response speed and Token cost, while enterprise administrators can configure model permissions by department, user group or member. Alibaba says Qwen4-Max supports continuous delivery, planning and optimization. Its reported evaluation results showed improvements over Qwen3.8-Max on professional Agent benchmarks and a Work360 average cost per task of 11.16 yuan, down from 26.58 yuan.
品玩 实时要闻 · 🔥 0 - Tsinghua's CoACT cuts coding-agent token use by 33%
Machine Intelligence reports that researchers from Tsinghua University proposed CoACT, an observation-compression method that preserves a coding agent’s next action rather than measuring similarity to the original observation. On 200 SWE-bench Verified instances with three agent models, CoACT reportedly reduced total token consumption by 33.0% on average, while average PASS@1 rose from 69.0% to 71.0% and interaction steps fell from 56.5 to 51.2. The study reports that CoACT can also be combined with AgentDiet, with estimated total cost falling to $25.88 while task success remained similar.
机器之心 · 🔥 0 - Blackwell fingerprints found in six Chinese model weights
Unemployment Labs researcher Will DePue found Blackwell-compatible numerical patterns in the published weights of six models released since April 2026 by DeepSeek, Moonshot AI, MiniMax and Ant Group, while earlier releases matched NVIDIA's Hopper generation. The analysis reproduced the initialization pattern for DeepSeek-V4-Flash and Kimi K3 on physical Blackwell hardware, providing evidence that the developers had access to such chips. The fingerprints do not establish where the chips were located, how access was obtained, whether export controls were violated, or whether Blackwell handled the entire training run.
Hacker News · AI(100+ 分) · 🔥 0 - aiXcoder open-sources 4B context compression model for coding agents
aiXcoder has open-sourced aiXcompact-4B, a 4B-parameter model that compresses coding-agent histories by retaining key requirements, condensing tool results, removing stale records and producing structured task summaries. The model supports private deployment on a single A100-40GB GPU. aiXcoder claims that, when connected to OpenCode, it matched DeepSeek-V4-Pro’s overall task-completion accuracy on Terminal-Bench, while cutting private-deployment inference costs by 78% for specified tasks; these benchmark and cost figures have not been independently verified.
智东西 · 🔥 0 - Recovering a Claude Code session that exceeded its 1M-token limit
A firsthand account describes a Claude Code session whose requests all failed with a 400 error: 930,665 message tokens plus a 128,000-token completion allowance exceeded the deepseek-v4-pro limit of 1,048,576 tokens. The author notes that /compact also returned 400, because it sends the full conversation to the model rather than truncating locally, and attributes the failure to three overlapping causes: the hard model limit, an auto-compact window set to the full 1M with no headroom, and /compact needing the full transcript. The suggested recovery order is to exit and relaunch with claude --resume and a reduced 65,536-token output allowance, then /compact immediately, and only then /clear if 400 persists, since session JSONL files and memory files survive. Prevention settings are an 800,000-token auto-compact window and a 65,536 output allowance, which take effect only for newly started processes and can be checked with /context. The session ran on Claude Code 2.1.308 with deepseek-v4-pro on 2026-11-02; the author notes context-management behavior may change across versions.
掘金 · 人工智能 · 🔥 0 - AREX-2 trains a 27B model with long-horizon reflection
The AREX research group released AREX-2, a model based on Qwen3.8-27B and trained with long-horizon reflection data for machine-learning and algorithm-programming tasks. The team reported scores of 81.8 on MLE-bench Lite, 70.7 on Frontier-CS, 84.0 on BrowseComp and 52.6 on HLE. It also said the training helped capabilities transfer to deep research and scientific problem-solving. The report presents long-horizon reflection as a trainable and verifiable approach to recursive self-improvement, which is the team’s research interpretation rather than established proof of that capability.
机器之心 · 🔥 0 - China report benchmarks LLM token-service performance
According to IT Home reports citing the joint publication, the China Academy of Information and Communications Technology's AI Research Institute, its Key Laboratory for Large Model and Software/Hardware Evaluation, and the China AI Industry Development Alliance released the 2026 LLM Token Service Performance Monitoring Report. Based on January-August monitoring data, it examines call success rate, throughput, time to first token, stability, price, context-window length and performance differences across official and third-party platforms. It reports that domestic daily token consumption reached about 500 trillion in June, while average domestic output throughput reached 78.23 tokens per second in August, compared with 38.25 for the 2025 average and 112.82 for typical foreign models.
IT之家 AI · 🔥 0 - DeepSeek-V4.1-Flash: 552B multimodal MoE with 1M context
DeepSeek published DeepSeek-V4.1-Flash on Hugging Face, a natively multimodal mixture-of-experts model with 552B backbone parameters and up to a 1M-token context, built on a Causal Encoder-Decoder design (20 encoder plus 20 decoder layers) that activates only 8B parameters per token in prefill and 16B in decode. It combines Compressed Sparse Attention 2 with FP4 main-KV caching (E2M1 format, one E4M3 scale per 16 channels) to cut the global KV cache to about 890 bytes per token — roughly a quarter of DeepSeek-V4-Flash — and offers a continuously controllable reasoning effort from 1 to 100, with weights under the MIT license. The vendor's own model card reports Terminal-Bench 2.1 at 90.6 (versus Opus-5.0 at 89.1 and GPT-5.6 Sol at 88.8), DeepSWE v1.1 at 74.2 and a Codeforces rating of 3471, but these figures are self-reported and not independently verified; the release also omits a Jinja chat template and ships a deepseek-recipe toolkit for prompt encoding instead.
DeepSeek · Hugging Face · 🔥 0 - ByteDance Seed paper ties DeepSeek long-context swings to KV-cache phase sensitivity
Per DoNews, ByteDance's Seed team posted an arXiv paper in late September 2026 arguing that "phase sensitivity" introduced by chunked KV-cache compression is the main driver of long-context retrieval fluctuations in DeepSeek-V4 series models. Covering open-weight models including DeepSeek-V4-Flash, V4-Pro and V4.1-Flash, the work says compression windows impose token phase coordinates, so the same information can be retrieved with accuracy differing by as much as 40 percentage points across phases — a periodic degradation that averaged benchmark scores tend to hide. The team calls for better cache-compression designs to improve long-context stability.
DoNews · 🔥 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.