Desk · Labs & reverse engineering

Research

Technical deep dives, reproducible tests, and tool evaluations.

Articles: 177
Topics: 20
Page: 12 / 23

Intelligence dossier infographic detailing DeepSeek V4-Pro's 14x pricing multiplier, 50 percent off-peak time-of-day billing schedule, and custom silicon capital expansion.

DeepSeek Upends AI Economics with V4 Pro Launch and Peak-Hour API Surcharges

DeepSeek launches DeepSeek-V4-Pro and introduces dynamic peak-hour API pricing, ending flat ultra-cheap rates while expanding custom silicon ambitions.

Aug 18 · 6 min

Data graphic: OpenCode Go's monthly usage caps per model on its $10/mo plan, with DeepSeek V4 Pro cut 75% to $15 and V4 Flash cut 50% to $30, while MiMo-V2.5, GLM-5.2 and Qwen3.7 Plus keep $60 and Grok and Kimi sit at $15. Working all peak hours, when DeepSeek rates double 01:00-04:00 and 06:00-10:00 UTC , roughly halves estimated monthly requests: about 65,000 to 32,500 on V4 Flash and 5,200 to 2,600 on V4 Pro.

stableResearch Opencode Go

OpenCode Go Slashes DeepSeek Quotas as AI API Reseller Economics Fracture

OpenCode Go cuts DeepSeek V4 quotas by up to 75% as peak API pricing and heavy token burn squeeze the margins of fixed-rate AI subscription aggregators.

Aug 18 · 6 min

Architecture diagram illustrating full VRAM residency for 27B models on a 24GB GPU, showing weight sizing, KV cache headroom, and memory bandwidth performance.

stableResearch 27b Llm Local Deployment

Local Deployment of 27B Large Language Models: The 24GB Hardware Blueprint

Learn how to run 27B LLMs on a 24GB GPU. Master VRAM sizing, GGUF/EXL2 quantization, and Ollama configs to achieve high-throughput local inference today.

Aug 18 · 7 min

Data graphic: Llama 3.3-70B needs 131.5 GB for its FP16/BF16 weights, which takes two 80GB H100s TP=2 unquantized, while Qwen 2.5-32B needs 60.5 GB and runs native BF16 on one H100; bars compare them with Gemma 2 27B 50.7 GB and Mistral Small 3 44.7 GB against 24GB and 80GB GPU lines. A benchmark table shows Qwen ahead on MATH 83.1% vs 77.0% and LiveCodeBench 51.2% vs 43.5% , and Llama ahead on IFEval 92.1% vs 79.5% and MMLU 86.0% vs 83.3% .

stableResearch Qwen 2 5 32b

Qwen 2.5-32B vs. Llama 3.3-70B: Enterprise Foundation Model Sizing, Benchmarks, and Local VRAM Economics

Compare Qwen 2.5-32B vs. Llama 3.3-70B on benchmarks, GPU VRAM sizing, and costs. Discover the best open-weight LLM architecture for enterprise AI deployments.

Aug 15 · 8 min

Technical dossier comparing Qwen2.5-Coder-32B against proprietary coding models across HumanEval+, SWE-bench Verified, VRAM requirements, and deployment profiles.

stableResearch Qwen2 5 Coder 32b

Open-Weight Coding at the Frontier: Evaluating Qwen2.5-Coder-32B for Enterprise Software Engineering

Compare Qwen2.5-Coder-32B with Claude 3.5 Sonnet and GPT-4o. Explore benchmark results, local hardware sizing, TCO analysis, and enterprise coding deployment.

Aug 15 · 6 min

Architectural infographic diagram detailing the total cost of ownership and routing logic between dedicated L40S GPU hosting and managed LLM APIs for 27B models.

stableResearch Llm Inference Economics

Enterprise Inference Economics: The Total Cost of Ownership Blueprint for 27B LLMs

Compare 27B LLM inference costs: self-hosted GPUs vs managed APIs. Explore break-even metrics, FP8 VRAM math, L40S benchmarks, and total cost of ownership.

Aug 15 · 8 min

Architectural comparison diagram showing latency, KV-cache footprint, and data routing flows between 8K vector RAG and 262K native long-context models.

stableResearch Qwen 3 8 27b

Qwen 3.8 27B vs. Vector RAG: Architectural Realities of the 262K Token Window

Compare Qwen 3.8 27B's 262K context window to vector RAG. Analyze KV cache VRAM limits, multi-hop reasoning decay, and hybrid enterprise AI architectures.

Aug 15 · 6 min

Architecture diagram illustrating local agent orchestration with Qwen 27B on vLLM, including schema validation, sandboxed execution, and security controls.

stableResearch Qwen 27b

Mastering Local Agent Orchestration: Production Tool-Calling with Qwen 27B and vLLM

Learn how to deploy autonomous agents locally with Qwen 27B and vLLM. Master tool-calling, grammar-constrained decoding, LangGraph, and enterprise guardrails.

Aug 15 · 6 min