Desk · Labs & reverse engineering

Research

Technical deep dives, reproducible tests, and tool evaluations.

Articles: 177
Topics: 20
Page: 13 / 23

Data graphic: Alibaba's open-weight Qwen3.8-27B scores 84.3% on OSWorld-Verified desktop OS automation, ahead of Qwen3.7-Plus at 73.3% and Claude Opus 4.6 Max at 72.7%, and beats Opus 4.6 Max on five of six benchmarks, including MathVision 94.6% vs 65.5% and SWE-bench Pro 61.7% vs 53.4% , losing only Terminal Bench 2.1 73.0% vs 78.2% . The 27B dense model has a 262,144-token native context 1M extended and 64 layers, 48 Gated DeltaNet plus 16 GQA.

Alibaba Releases Qwen3.8-27B: Open-Weight Hybrid Model Sets Frontier Agentic Benchmarks

Alibaba launches Qwen3.8-27B, an open-weight hybrid model combining linear attention with GQA to deliver frontier coding, OS control, and agent performance.

Aug 15 · 6 min

Data graphic: Qwen 3.8 27B leads OCRBench dense multi-layout OCR with 885 points, ahead of Claude 3.5 Sonnet 788 , Llama 3.2 90B Vision 762 , GPT-4o 736 and Llama 3.2 11B Vision 684 , 149 points above GPT-4o; it also leads DocVQA 96.4% and MathVista 74.8% but trails Claude 3.5 Sonnet on ChartQA 89.5% vs 90.8% and MMMU-Pro 51.1% vs 54.7% , and at FP8 it needs about 27 GB of VRAM and runs at 85–120 tokens per second. Verdict: self-host above 100,000 pages a month or under an air-gap mandate, and keep frontier APIs for intermittent or open-domain work.

stableResearch Qwen 3 8 27b

Open-Weights Frontier: Qwen 3.8 27B Sets New Multimodal Benchmarks for Enterprise Vision Workflows

Explore Qwen 3.8 27B multimodal benchmarks, architecture, and deployment strategies. Learn how open-weights vision AI rivals proprietary enterprise APIs.

Aug 15 · 7 min

Data graphic: serving Qwen 27B on one 80 GB H100 or A100, where FP8 fits about 48 streams at 8k context against about 11 at FP16, because the weights shrink from 54.0 GB at FP16 to 27.0 GB at FP8 14.85 GB at INT4 . At concurrency 64 with FP8, SGLang does 1,980 tok/sec, TensorRT-LLM 1,920 and vLLM 1,650, and the verdict picks SGLang for agentic and JSON-heavy work, TensorRT-LLM for strict latency SLAs and vLLM for fast cloud-native deploys.

stableResearch Qwen 27b Inference

High-Concurrency Inference for Qwen 27B: Sizing, Benchmarks, and Production Engine Architecture

Master high-concurrency Qwen 27B inference. Compare vLLM, SGLang, and TensorRT-LLM benchmarks, GPU memory sizing, FP8 quantization, and hardware topologies.

Aug 15 · 7 min

Data graphic: fine-tuning Qwen 2.5 32B with Unsloth and 4-bit QLoRA takes 20.8GB of VRAM for the whole training pipeline, against upwards of 120GB for standard 16-bit fine-tuning, leaving 3.2 GB free on a 24 GB GPU. The budget splits into about 14.8 GB of 4-bit base weights, 2.8 GB of activations and KV cache, 2.2 GB of CUDA context, 0.65 GB of optimizer states and 0.35 GB of LoRA adapters, alongside hyperparameter targets such as LoRA rank 16 or 32 and a learning rate of 1e-4 to 2e-4.

stableResearch Unsloth Fine Tuning

Fine-Tuning 27B–32B LLMs on a Single 24GB GPU: Unsloth and 4-Bit QLoRA Production Guide

Learn how to fine-tune 27B–32B LLMs like Qwen 2.5 on a single 24GB GPU using Unsloth and 4-bit QLoRA with full code, memory optimization, and serving guides.

Aug 15 · 7 min

Technical dossier infographic displaying GLM-5.3 cybersecurity capabilities, benchmark scores including 84.5 percent CyberGym and 54.4 percent ExploitBench, vulnerability discovery volume, and enterprise verification bottleneck metrics.

stableResearch Glm 5 3

GLM 5.3 Cybersecurity Benchmarks and Enterprise Security: Frontier Capabilities Meet Operational Realities

Z.ai's GLM-5.3 achieves an 84.5% vulnerability discovery rate on CyberGym, outperforming frontier AI models while creating new SecOps verification challenges.

Aug 15 · 6 min

Technical dossier of xAI Grok 4.6 showing core performance metrics, 500,000-token context window, pricing structure, and comparative benchmark breakdown.

stableResearch Xai Grok 4 6

xAI Releases Grok 4.6 With 500k Context Window, Challenging Frontier Pricing While Coding Deficits Persist

xAI launches Grok 4.6 with a 500k context window and aggressive pricing, matching GPT-5.6 Sol in knowledge work while trailing in autonomous coding benchmarks.

Aug 15 · 6 min

Technical architecture diagram of Anthropic Claude Opus 5 and locked-rate Sonnet two-tier enterprise deployment pipeline, showing benchmarks and cost metrics.

stableResearch Anthropic

Anthropic Launches Claude Opus 5 and Freezes Sonnet Pricing to Reshape Enterprise AI Economics

Anthropic launches Claude Opus 5 with 1M context alongside a permanent price freeze on Sonnet, cutting enterprise AI costs and setting new SWE-bench records.

Aug 15 · 6 min

Architecture diagram of OpenAI GPT-5.6 Sol dynamic test-time compute routing and ultrafast wafer-scale hardware acceleration tier.

stableResearch Openai Gpt 5 6 Sol

OpenAI GPT-5.6 Sol Reasoning Slider Hardware Preview: Silicon Co-Design and the High-Stakes Shift to Dynamic Test-Time Compute

OpenAI previews GPT-5.6 Sol, combining dynamic test-time compute via a Reasoning Slider with ultrafast silicon delivering up to 750 TPS reasoning speeds.

Aug 15 · 8 min