Deepseek-V4-Flash
AI overview
Sign in and the AI will write an overview from our coverage.
Headlines · 11
- LLMs Show “Thinking Inertia” When Told to Answer Only
Researchers from Tsinghua University, the University of Oxford and Stanford University tested six models from the Qwen3, DeepSeek, GPT and Gemini families across six benchmarks. They found that disabling thinking mode and asking for answers without reasoning reduced, but did not eliminate, visible calculations or derivations. Explicit reasoning was most common on open-ended questions, while yes-or-no questions prompted more direct answers. The study introduced three metrics to assess answer-only compliance, relevance of preceding text and explicit reasoning separately. The paper, code and trajectory data are public. The source reported that the paper was accepted at NeurIPS 2026, and a separate community member ran a small-scale independent test that also observed reasoning after thinking was disabled.
新智元 · 🔥 0 - Coding-Agent Harness Changes Raise Costs Without Reliable Gains
Shandong University researchers Liu and Han compared Claude Code, OpenCode, and mini-SWE-agent while holding the language model constant. Most performance differences were similar to rerun noise, while the harnesses consumed two to three times as many tokens as the leanest option; OpenCode showed a significant deficit on the larger task pool. A small cross-check on Terminal-Bench 2.1 showed the same pattern.
The Decoder · 🔥 0 - Snowflake's Cortex AI Gateway enters public preview on AWS
Snowflake has moved its Cortex AI Gateway into public preview in AWS commercial regions, according to InfoQ's relay of a Snowflake blog post. The gateway acts as a central control plane for enterprise AI, offering governed model inference through a single endpoint, dynamic model routing, MCP tool governance, observability, guardrails and cost controls. It implements the Chat Completions API for OpenAI, Grok, GLM, Gemini, Llama, Mistral and DeepSeek models and the Messages API for Claude, with access managed through Snowflake's existing role-based access control.
InfoQ 中文 AI&大模型 · 🔥 0 - Ant Digital launches LingDT-Decision-1.0 decision models
Ant Digital Technology released LingDT-Decision-1.0 in 4B and 27B versions for structured enterprise tasks such as intent recognition, classification, tool selection, routing, and risk assessment. The models add native multi-select and dependency-aware DAG sequencing, and are available through the DTMaaS platform with a limited free trial. Ant Digital reported lower token use, faster processing, and leading scores across 23 open-source benchmarks, but the supplied reports contain no independent validation of those performance claims.
InfoQ 中文 AI&大模型 · 🔥 0 - MemT 1.0 Opens the Source for Memory-Native AI Models
MemTensor released and open-sourced MemT 1.0, a family of memory-native foundation models comprising MemT-1.0-Nano and MemT-1.0-Flash. The architecture uses Gated DeltaNet-2 and Mixture-of-Storage to maintain updating internal memory states that participate in generation, with author-reported gains in long-history inference efficiency, storage reduction, and agentic task success. External text-memory systems still retain an advantage for precise recall from long histories, according to the reports.
Hugging Face · Papers · 🔥 4 - vLLM brings adaptive verification to draft models without a confidence head
vLLM's blog says Adaptive Verification now supports all draft-model-based speculator types, not just DSpark drafters that ship with a confidence head. For drafters lacking one, such as EAGLE, MTP, DFlash and some DSpark checkpoints, vLLM uses a lightweight logistic regression that maps the log-odds of the drafter's most likely token to the target model's acceptance probability, with one shared slope plus one intercept per draft position (K+1 parameters) fitted online from the accept/reject decisions produced during normal verification. The feature landed in PR #52228 and is enabled with enable_adaptive_verification; on DeepSeek-V4-Flash with DSpark the estimator beat the trained confidence head on AUC and log-loss but had worse ECE, and the two were nearly tied on calibration once the first draft position was excluded. Stated limitations include requiring a target attention backend that supports variable-length verification, and smaller gains for MTP and EAGLE than for DFlash and DSpark.
vLLM Blog · 🔥 0 - vLLM open-sources bitwise-consistent RL training kernels and an agent skill
A vLLM blog post argues that inference-side optimizations in high-performance RL — low-precision compute, low-precision KV cache, kernel and communication tuning — shift the numerics, so the training engine assigns different probabilities to the same token than the engine that sampled it; on DeepSeek-V4-Flash the high-performance baseline's train-inference KL rose from 2e-3 to 29.9 as reward fell back. The proposed fix has the training forward call vLLM's inference kernels directly, modified so rollout and training produce bitwise-identical logprobs for every sampled token (diff = 0). The authors report diff of 0 at every one of 100 DeepSeek-V4-Flash steps at 91.3% of the baseline's rollout throughput, and say that on Nemotron-3.5 (a Mamba hybrid with NVFP4) the consistent run's SWE-bench Pro score held near 0.6 while the baseline's fell from 0.56 to 0.22. The workflow is packaged as an agent skill and released with recipes and code, and the team says DeepSeek-V4.1 has reached 0 diff with performance work still in progress.
vLLM Blog · 🔥 0 - Tsinghua team's ActionOPD distills LLM agents at the action level, not tokens
A Tsinghua University research team posted ActionOPD on arXiv (2610.19820, submitted 14 Oct), an action-level on-policy distillation method that samples actions from both the student and a black-box teacher at states the student encounters, rewards student actions by semantic agreement with teacher actions, and optimizes the policy without needing teacher token probabilities or a shared model family or tokenizer. According to a 36Kr write-up by Academic Headline, the authors report that Qwen3-4B and Qwen3-8B students trained with two teachers (Qwen3.7-Flash and DeepSeek-V4-Flash) gained 3.3-6.9 percentage points on average across four out-of-distribution agent benchmarks. On τ²-bench, ActionOPD reached 40.0% pass¹ versus 37.5% for offline distillation and 39.6% for token-level OPD, but its 14.2% pass⁴ trailed their 19.2% and 17.5%; the paper also explores Self-ActionOPD and lists limitations such as semantic agreement not guaranteeing correct decisions.
Hugging Face · Papers · 🔥 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 - Blog SGLang and Miles on NVIDIA Vera Rubin We had early access to two NVIDIA Vera Rubin nodes (8 GPUs) and used them to bring SGLang and Miles up on the platform: SGLang for inference and Miles for RL training. SGLang & Miles Team October 9, 2026
On this page Summary Inference: Serving Kimi K3 NVFP4 Attention MoE and communication kernels KDA RL training: Miles on NVIDIA Vera Rubin Agentic RL on the Vera CPU What's Next Acknowledgements Summary We had early access to two NVIDIA Vera Rubin nodes (8 GPUs) and used them to bring SGLang and Miles up on the platform: SGLang for inference and Miles for RL training. On the inference side, we tuned SGLang for Kimi K3 NVFP4: this post covers what we changed in the attention, MoE/communication and KDA kernels, and how much each change bought. On the training side, Miles, which uses SG
LMSYS Blog · 🔥 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-Flash experienceAsk about Deepseek-V4-Flash
Nobody has shared their experience with Deepseek-V4-Flash yet.