DeepSeek-V4-Pro

Headlines · 4

1. 5d
   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
2. Oct 23
   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
3. Oct 12
   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
4. Oct 9
   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