Desk · Labs & reverse engineering

Research

Technical deep dives, reproducible tests, and tool evaluations.

Articles: 177
Topics: 20
Page: 2 / 23

Data graphic: slope chart of vLLM startup on H200, before v0.30 versus v0.30: engine init from 28.9s to 8.2s 20.7s saved, 3.5x faster and CUDA graph capture from 12s to 2s 10s saved .

vLLM v0.30.0 Upgrade Notes: Fast Start, HiSparse, New Models and the Breaking Changes

vLLM v0.30.0 adds DeepSeek-V4.1-Flash, GLM-5.3-Flash, Fast Start and HiSparse, and changes defaults that can break an upgrade. What to check.

Oct 2 · 4 min

Data graphic: MiniMax M3.1 Flash, a multimodal coding model that always thinks before it answers, shown with a huge "1M" for its 1-million-token context window. Two smaller stats give its 5 effort levels low to max, default max and a 66.25% score 53/80 points on KingBench 3, an independent 8-task test.

watchResearch Minimax M3 1 Flash

MiniMax M3.1 Flash: 1M-Context Coding Model That Always Thinks

MiniMax's M3.1 Flash preview brings a 1M-token context, image and video input and five effort levels to MiniMax Code. What it does, and how it tests.

Oct 2 · 4 min

Data graphic: Xing4.0-29B-A4B, Qwen3.6-35B-A3B and Gemma4-26B-A4B on four vendor-reported benchmarks. Xing4.0 scores 75.00 on SWE-bench Verified against Qwen3.6's 76.00, and leads Terminal-Bench 2.1 with 57.50 to Qwen's 51.50, a 6-point gap.

watchResearch Open Weight Coding Models

Xing4.0-29B-A4B: China Telecom's Open Coding Model Claims SWE-bench 75 on One GPU

China Telecom's Apache 2.0 MoE, trained on Huawei Ascend, scores 75.00 on SWE-bench Verified in its own table, a point behind Qwen3.6, and its 4-bit file is ~18 GB.

Oct 2 · 6 min

Data graphic: Copilot drops six models. GPT-5.5, GPT-5.4, GPT-5.4 mini, GPT-5 mini, Gemini 3.7 Flash and Grok 4.5 are struck through, each with its replacement, beside a huge 19 for the 19 Oct 2026 cut-off.

watchResearch Github Copilot

GitHub Copilot Drops Six Models on Oct 19: What Replaces GPT-5.5, GPT-5.4 and Grok 4.5

On 19 Oct GitHub Copilot drops GPT-5.5, GPT-5.4, two mini models, Gemini 3.7 Flash and Grok 4.5. Here is each replacement, what it costs and what admins must do.

Oct 2 · 7 min

Data graphic: Claude Sonnet 4.5 retires on Nov 30, 2026, after 61 days of notice from Sep 30. Sonnet 5.5 costs $2 / $10 per million tokens, down from $3 / $15, uses about 30% more tokens for the same text, and returns a 400 error on request shapes Sonnet 4.5 accepted.

elevatedResearch Anthropic

Claude Sonnet 4.5 Retires Nov 30: Six Requests That Return a 400 on Sonnet 5.5

Sonnet 4.5 stops answering on Nov 30, 2026. Sonnet 5.5 is a third cheaper per token, but prefill, forced tool use and thinking budgets now return 400 errors.

Oct 2 · 8 min

Data graphic: Gemini 4 Argon scores 52.6 on the Artificial Analysis Intelligence Index, level with GPT-6 Astra 52.7 and behind Claude Opus 5.5 57.6 . Cost per task: Argon $1.99 at intro price, about $3.98 at full price, Astra $3.26, Opus 5.5 $5.98. The leaked 88% DeepSWE score sits beside the real 77.9%, and the $2/$10 intro price beside the later $4/$20.

watchResearch Gemini 4 Argon

Gemini 4 Argon vs GPT-6 Astra vs Claude Opus 5.5: Is Google's New Model Actually Better?

The leaked 'Gemini 4 Pro' was billed as an Astra and Opus killer. The real Gemini 4 Argon wins Google's tests, ties Astra on neutral ones, and is cheap for now.

Oct 1 · 6 min

Data graphic: comparison of three Chinese Flash MoE models. Ling 3.1 Flash is the largest at 560B total and ~25B active parameters with no DeepSWE 1.1 score published; GLM 5.3 Flash 320B total, 18B active, MIT, $0.15/$0.50 per 1M tokens scores 63.4; Qwen 3.8 Flash Next 125B total, 6B active, $0.16/$0.47 scores 58.7.

stableResearch Ling 3 1 Flash

Ling 3.1 Flash vs GLM 5.3 Flash vs Qwen 3.8 Flash Next: Three Bets on Cheap Agent Models

Ant Group's 560B Ling 3.1 Flash lands against GLM 5.3 Flash and Qwen 3.8 Flash Next. Specs, licences, prices, and the benchmarks Ling has not published yet.

Sep 30 · 6 min

Data graphic: Ant Group's Ling-3.1-flash posts a vendor-reported 87.90 on CyberGym. On SWE-Pro it scores 65.39, behind Claude Opus 5 79.20 but ahead of Kimi K3 64.84 , GLM 5.3 flash 63.06 and DeepSeek-V4.1-Flash 56.77 . It grows from 124B to about 560B total parameters and from 5.1B to about 25B active, and ships as a free two-week trial capped at 256K context, with no safety evals and no weights yet.

elevatedResearch Ling 3 1 Flash

Ling-3.1-flash Is Ant Group's Best Flash Model Yet, and It Scores 87.9 on CyberGym

Ant's 560B-parameter Ling-3.1-flash beats rival flash models on SWE-Pro and posts 87.90 on CyberGym, but it ships with no safety evals and no weights yet.

Sep 30 · 6 min