DeepSeek-V4 1M Context vs RAG: Why Retrieval Stays
DeepSeek-V4 markets a 1M token window for agents, but no independent benchmarks verify depth recall. CEO-Bench and ProGraph research show structured memory outperforms raw.
The archive · Page 3 of 6
Models & Research
Where architecture, training tricks, and eval methodology meet the marketing layer — separating durable progress in foundation models from leaderboard theater that quietly falls apart under load.
The archive
49–72 of 139 articles · Newest first
Qwen-Image-3.0 Does Not Exist: Why Self-Hosting Image Models Is Premature
No primary source confirms Qwen-Image-3.0 exists. Independent benchmarks show open-weight models fail 85% of precise image tasks. Self-hosting branded templates remains risky.
Kimi K3: 2.8T Parameters, MoE Routing, and Self-Hosting Reality
Kimi K3 pairs 2.8T parameters with 16-of-896 expert routing and a 1M context window. Hosted API is live, but full weights arrive July 27, 2026. Self-hosting requires.
Kimi K3 vs Qwen3.8 Max: Routing Strategy for July 2026
Kimi K3 offers concrete pricing for production bake-offs while Qwen3.8 Max remains a shadow evaluation. Compare both against Fable 5, GPT-5.6 Sol, and GLM-5.2 by workload.
Qwen3.8 Max Release Audit: API, Open Weights, and the License Catch
Qwen3.8 Max now has a stable API, $2/$6 pricing, 1M context, and downloadable 2.4T weights. The release is real, but the API and open checkpoint are not the same product.
HuggingFace 100x Inference: Generalizable vs Platform-Locked Optimizations
HuggingFace claims 100x inference speedup, but independent benchmarks show its TGI engine trails vLLM by 24x on throughput. This article decomposes the claim to separate.
Kimi K3 Code Arena Rank: Self-Hosting Cost Math for Coding Agents
Kimi K3 is frontier-class, but its 2.8T MXFP4 checkpoint needs 64-plus accelerators. The real deployment math reveals what an arena rank cannot.
Can Tool-Adaptive LLM Rerankers Improve RAG Without Always Calling Tools?
TALRanker folds the tool-call decision into the reranker's scoring policy, turning tool latency from a fixed per-query tax into a budget the model spends only when uncertain.
Does Speculative Decoding with Progressive Tree Drafting Cut LLM Latency?
Progressive Tree Drafting grows draft tokens as a pruned tree to claim a 2× speedup, but the gain hinges on verifier acceptance and shrinks on code and reasoning.
Analytic Inference Cuts Bayesian Deep Ensemble Serving Cost, But Leaves Training as the Bottleneck
A new arXiv preprint replaces sampling-based averaging in Bayesian deep ensembles with closed-form Bayesian aggregation, cutting per-query inference cost and shifting the.
Tree-of-Thoughts Improves Text-to-Image Prompting by Reasoning Over Hypotheses, Not Pixels
A new arXiv paper ports Tree-of-Thoughts prompting to text-to-image in-context learning and reports CoBSAT gains, but reasoning runs on prompt hypotheses, not image states.
FourierQK's spectral Q/K filter cuts TinyShakespeare loss by 79%, but long-context proof is missing
FourierQK filters query and key projections before attention, cutting TinyShakespeare character-level loss by 79%, but word-level, retrieval and long-context tests are absent.
Tencent Hunyuan 3's Agent Push Has No Public DeepSeek or Qwen Benchmarks Yet
Tencent's Hy3 is billed as a 295B agent model with a 21B active token footprint, but public materials omit benchmark tables and independent DeepSeek or Qwen comparisons.
When Does Memory, Not Compute, Decide Who Can Profitably Serve LLMs?
A July 2026 arXiv paper argues that scarce HBM and DRAM bandwidth, not raw compute, will determine which labs and providers can profitably serve large language models through.
Can We Trust LLM Logic? A Graph-Based Stress Test Finds Three Failure Modes
A July 2026 arXiv paper shows Self-Consistency voting can hide contradictory reasoning, and GraphEVAL's graph-based coherence metrics catch flawed paths output checks miss.
LongCat-2.0 Hits Claude Opus 4.6 Class on Agents From a 50K-GPU Cluster
LongCat-2.0 is a 1.6T MoE trained on a 50,000-card domestic cluster without NVIDIA silicon, posting agent scores near Claude Opus 4.6 yet CUDA serving stacks still dominate.
When Do Time Series Foundation Models Pay Off? The Break-Even Threshold
A July 2026 arXiv break-even study finds pretrained time series foundation models win unconditionally on half of 30 datasets, but lose to ARIMA or XGBoost on a fifth at any.
VLA Grounder Tests Language Conditioning to Optimize Black-Box Vision-Action Models
VLA Grounder shows frozen vision-language-action models improve when language is an optimizable input. Success rates increased from 12.6% to 38.7% for pi0 and 24.4% to 63.9%.
Black-Box LLM Architecture Inference: What API Restrictions Reveal About Hidden Model Structure
NightVision recovers transformer hidden dimension to within 23% error using only single logprob output and timing, proving API restrictions meant to protect model IP instead.
InduceKV Tests Continual Learning for Multimodal LLMs Without Expanding Cache
InduceKV keeps multimodal LLM adaptation under a fixed KV memory budget, outperforming replay and PEFT while trading selection complexity for serving simplicity.
Sonnet 5 vs GPT-5.5: Pricing, Benchmarks, and the Switching Math
Sonnet 5 undercuts GPT-5.5 by 60% on input tokens and leads coding benchmarks, but a new tokenizer inflates effective costs and Anthropic did not publish GPQA scores.
How LLMs Fuse Conflicting Facts: Single-Source vs Multi-Source Truth
LLMs beat classic truth-discovery on conflicting facts, but the same reasoning treats repeated low-credibility claims as corroboration. More sources do not mean more truth.
Linear Transformers Get a Learnable Kernel: Does Flexformer Change the Efficiency Tradeoff?
Flexformer makes linear attention's kernel learnable by training spectral frequencies, but the abstract offers no perplexity or accuracy numbers to back its gains.
Huawei Ships CUDA-Free AI Compute On-Device, but Ascend Quantization Accuracy Is Unverified
Huawei ships CUDA-free AI compute on domestic silicon today, but specific OpenPangu quantization accuracy claims on Ascend NPUs lack any readable primary source.
