Kimi K3.1 Pro(kimi.ai)
47 points by nekofneko | 5 comments
Moonshot AI blog post announcing Kimi K3.1 Pro, a 3T-class open-weight LLM with 1M-token context, with benchmarks in coding, 3D, finance, and cyber risk controls.
#benchmarks on Hacker News
60 stories tagged benchmarks, newest first.
Gemma-4-26B-A4B with 39-69% less thinking at same accuracy(bottlecapai.com)
103 points by armcat | 19 comments
BottleCap AI blog post announcing ThinkingCap-Gemma-4-26B-A4B, a fine-tune cutting thinking tokens by 39% on average with near-unchanged benchmark accuracy.
Remove pgrust: the entry is cheating (per-query kernels and result caching)(github.com)
166 points by f311a | 104 comments
A merged ClickBench pull request removing the pgrust entry for using hand-written per-query kernels and result caching that violate the benchmark rules.
LLM agents play Pokemon showdown(olivergrabner.com)
25 points by grabnero | 14 comments
A research write-up on a multi-agent environment where LLM teammates coordinate via chat in Pokémon Showdown 2v2 battles, comparing win rates across talk modes.
LabRobotBench v0.1: Benchmarking wet lab robotics(gamowlabs.com)
11 points by wardbradt | 1 comments
Gamow Labs research post introducing LabRobotBench v0.1, a benchmark for AI agents operating wet lab robotics in simulated lab environments.
Benchmarking Bot Detection Systems Against Modern AI Agents in 2026(research.poh.org)
37 points by mdahardy | 3 comments
Proof of Human research post benchmarking bot detection systems like reCAPTCHA v3, hCaptcha and Fingerprint Pro against stealth browsers and AI agents, finding 9% to 99% detection rates.
Putting Linux on my RTX Spark laptop because I don't have a Microsoft account(josie.lol)
234 points by pfeifferj | 120 comments
A personal blog post describing installing Arch Linux on an ASUS ProArt P16 laptop with an NVIDIA RTX Spark ARM64 chip, with setup patches and benchmarks against Windows.
Superlite: A Faster Alternatie to Superpowers(contextfirst.blog)
131 points by iheatuchriswogu | 61 comments
Blog post introducing Superlite, an autonomous build workflow for Claude Code that executes a spec with fewer permission prompts, covering usage and benchmarks.
![Screenshot of youtube.com: RuneScape Bench: MMOs for Agents [video]](https://hn.urlbox.com/shots/renders/2026/10/19/bbe86289-5380-459d-97f6-a6c945ddde1c.jpg)
RuneScape Bench: MMOs for Agents [video](youtube.com)
224 points by evakhoury | 74 comments
Recurse Center YouTube talk by Max Bittker presenting RuneScapeBench, a benchmark using a RuneScape server to evaluate AI agents in an MMO.
A look at the math that OpenAI claims computes faster Fourier transforms(argmin.net)
204 points by mochomocha | 129 comments
Ben Recht's argmin blog post critiquing OpenAI's math repository claim of a faster FFT algorithm, arguing it is more an exploit than a breakthrough.
WSL3 Performance is about 5-60% faster than WSL2 depending on the workload(tonym.us)
218 points by tonymet | 195 comments
Blog post benchmarking WSL 2.6.3 against WSL 3.0.2 across syscalls, memory bandwidth, and Go compile workloads, comparing Linux kernels 6.6 and 6.18.
Write Like It's 1866: LLMs Relearn Telegraphese(fiveminutesforward.com)
79 points by Theory42 | 51 comments
Blog post introducing the Telegraph Test benchmark, showing that instructing LLMs to answer in telegraph-style cablese cuts output tokens by 40–49% while preserving information.
Jev-Driven SRE Diagnosis: What Worked and What Failed(sregym.com)
32 points by matt_d | 11 comments
SREGym research blog post describing a Jev-driven pipeline for SRE diagnosis that picks where to investigate and which cluster evidence supports a diagnosis.
How Fast is Python 3.15?(blog.miguelgrinberg.com)
59 points by Qem | 39 comments
Miguel Grinberg's blog post benchmarking Python 3.15 (rc3) performance against earlier interpreters back to Python 3.10.
Benchmarking retrieval for agents on messy real-world company knowledge(kapa.ai)
22 points by emil_sorensen | 2 comments
Kapa blog post introducing Company Knowledge Bench, a benchmark for evaluating agent retrieval over messy real-world company knowledge.
Show HN: Open-source model routing for coding agents at Astra-level performance
99 points by adchurch | 32 comments
Show HN post announcing Weave Router 2.0, an open-source model router for coding agents that switches between LLMs to match GPT-6 Astra performance at lower cost.
Show HN: Jevstiller – Distill Jev into a local model, with a disagreement bound(jevstiller.pages.dev)
54 points by tgluck | 8 comments
Documentation post for Jevstiller, a tool that distills the Jev classification API into a local model, explaining its 98% agreement rate and disagreement guarantee.
TabPFN and TabICL vs. tuned XGBoost: the model that doesn't train won 14/14(efraingaray.com)
21 points by EfrainGaray | 9 comments
Blog post benchmarking TabPFN and TabICL foundation models against tuned XGBoost on fourteen tabular datasets, finding the no-training models won all fourteen.
Postgres SELECT DISTINCT Does Not Scale(dbos.dev)
83 points by KraftyOne | 26 comments
A DBOS blog post by Peter Kraft explaining why Postgres SELECT DISTINCT queries perform poorly at scale.
Best LLM for every budget, updated daily(bestmodelforyourbudget.terrydjony.com)
127 points by terryds | 80 comments
Interactive chart plotting LLMs on the Artificial Analysis Intelligence Index against blended API price, highlighting the value frontier with filters.
Show HN: JevBench, a reproducible benchmark for typed decision models(benchmarkheaven.com)
119 points by florianstandhar | 32 comments
Benchmark Heaven's JevBench v1.3.0 leaderboard page ranking 52 Jev-class typed decision models on intelligence, calibration, speed and cost across 534 decisions.
DeepSeek v4.1 Flash(twitter.com)
919 points by Liwink | 505 comments
DeepSeek's X post announcing DeepSeek-V4.1-Flash, a small model with native visual understanding, shown with benchmark comparisons against rival models.
I trained a small transformer in 1.5hrs and it beats many LLMs(mvakde.github.io)
578 points by porridgeraisin | 149 comments
Personal blog post describing a small transformer trained from scratch in 1.5 hours on an RTX 5090 that scores 44% on ARC-AGI-1 for 67 cents.
Why your local LLM feels dumber than it is(forum.level1techs.com)
417 points by felineflock | 171 comments
A Level1Techs forum post presenting technical experiments on how quantization and implementation differences make locally run LLMs perform worse than reference versions.
Assembly Hall of Shame(github.com)
403 points by piotrgrabowski | 99 comments
GitHub repository by xoreaxeaxeax collecting x86 assembly instructions and code patterns that perform pathologically slowly, such as clflush, cpuid, denormals and fdiv.
Evidence of inconsistencies in evaluation process and selection of winners(kaggle.com)
446 points by twerkmeister | 276 comments
A Kaggle discussion post alleging inconsistencies in the judging and winner selection of the 'Measuring Progress Toward AGI' competition, citing the Medley Bench grand prize winner.
GPT-5.5 hallucinates 3x more than MIT-licensed GLM-5.2(arrowtsx.dev)
474 points by oshrimpton | 235 comments
Blog post arguing that bigger AI models are not better, comparing benchmark and hallucination rates of GLM-5.2, GPT-5.5, Opus 4.8, DeepSeek V4 and others.
Disagreement among frontier LLMs on real-world fact-checks(lenz.io)
457 points by kostaj | 314 comments
Lenz research report finding that five frontier LLMs disagree on the verdict for 67% of 1,000 real-world fact-check claims.
MacBook Neo Deep Dive: Benchmarks, Wafer Economics, and the 8GB Gamble(jdhodges.com)
319 points by tosh | 373 comments
A personal blog post analyzing Apple's MacBook Neo, covering benchmarks, chip wafer economics, and its 8GB RAM configuration.
Computer Use is 45x more expensive than structured APIs(reflex.dev)
458 points by palashawas | 255 comments
Reflex blog post benchmarking AI computer use against auto-generated API endpoints on an admin panel, finding computer use far more costly in steps and tokens.
Ripgrep is faster than grep, ag, git grep, ucg, pt, sift (2016)(burntsushi.net)
339 points by jxmorris12 | 143 comments
Andrew Gallant's 2016 blog post introducing ripgrep, a Rust command-line search tool, with 25 benchmarks comparing it against grep, ag, git grep, ucg, pt, and sift.
Big data on the cheapest MacBook(duckdb.org)
360 points by bcye | 283 comments
DuckDB blog post benchmarking database workloads on Apple's entry-level MacBook Neo to test its suitability for big data analysis.
What Claude Code chooses(amplifying.ai)
565 points by tin7in | 210 comments
A study by Amplifying that ran Claude Code against real repos 2,430 times to analyze which tools and libraries it chooses, finding it tends to build custom solutions.
SkillsBench: Benchmarking how well agent skills work across diverse tasks(arxiv.org)
333 points by mustaphah | 142 comments
arXiv paper presenting SkillsBench, a benchmark of 86 tasks across 11 domains measuring whether structured Agent Skills improve LLM agent performance.
Improving 15 LLMs at Coding in One Afternoon. Only the Harness Changed(blog.can.ac)
776 points by kachapopopow | 278 comments
Blog post by Can Bölük arguing that changing only the edit tool in a coding agent harness improved coding performance across 15 LLMs.
Frontier AI agents violate ethical constraints 30–50% of time, pressured by KPIs(arxiv.org)
512 points by tiny-automates | 329 comments
arXiv paper introducing a 40-scenario benchmark measuring how autonomous AI agents violate ethical, legal, or safety constraints when pressured by KPI-driven outcomes.
Python numbers every programmer should know(mkennedy.codes)
382 points by WoodenChair | 157 comments
Michael Kennedy's blog post listing Python performance and memory numbers, such as list operations, file I/O, object sizes, and web framework speed.
100k TPS over a billion rows: the unreasonable effectiveness of SQLite(andersmurphy.com)
344 points by speckx | 124 comments
Blog post by Anders Murphy benchmarking SQLite at 100k transactions per second over a billion rows, arguing its embedded single-writer design helps it scale.
So you wanna build a local RAG?(blog.yakkomajuri.com)
367 points by pedriquepacheco | 97 comments
Blog post describing how the Skald team built a fully local, self-hosted RAG stack using open-source components, with benchmarks versus proprietary APIs.
Tongyi DeepResearch – open-source 30B MoE Model that rivals OpenAI DeepResearch(tongyi-agent.github.io)
340 points by meander_water | 136 comments
Blog post from Alibaba's Tongyi Lab introducing Tongyi DeepResearch, an open-source 30B MoE agent model, with benchmark results against OpenAI DeepResearch and others.
Kafka is Fast – I'll use Postgres(topicpartition.io)
500 points by enether | 359 comments
Blog post benchmarking Postgres as a pub-sub and queue system compared to Kafka, arguing for simpler, boring technology choices.
Python 3.14 is here. How fast is it?(blog.miguelgrinberg.com)
695 points by pjmlp | 505 comments
Miguel Grinberg's blog post benchmarking the newly released Python 3.14 against earlier Python versions to measure performance gains.
Vibechart(vibechart.net)
801 points by datadrivenangel | 173 comments
A satirical dictionary-style page defining "vibechart" as charting what you want to see, illustrated with misleading bar charts from OpenAI's GPT-5 release.
Show HN: Factorio Learning Environment – Agents Build Factories(jackhopkins.github.io)
695 points by noddybear | 200 comments
Project page for the Factorio Learning Environment, an open-source framework and benchmark for evaluating LLM agents building factories in Factorio.
Performance of the Python 3.14 tail-call interpreter(blog.nelhage.com)
557 points by signa11 | 160 comments
Blog post by Nelson Elhage examining the performance gains of CPython 3.14's new tail-call bytecode interpreter and reassessing its reported 10-15% speedup.
An analysis of DeepSeek's R1-Zero and R1(arcprize.org)
703 points by meetpateltech | 257 comments
ARC Prize blog post by Mike Knoop analyzing DeepSeek's R1-Zero and R1 models, arguing R1-Zero is more important, with ARC-AGI benchmark results.
OpenAI O3 breakthrough high score on ARC-AGI-PUB(arcprize.org)
1578 points by maurycy | 1581 comments
ARC Prize blog post by François Chollet reporting OpenAI's o3 scoring 75.7% (and 87.5% at high compute) on the ARC-AGI semi-private evaluation set.
Something weird is happening with LLMs and chess(dynomight.substack.com)
626 points by crescit_eundo | 424 comments
A Dynomight newsletter post examining how well large language models play chess and why results vary strangely across models.
M4 Mac mini's efficiency(jeffgeerling.com)
518 points by marinesebastian | 515 comments
Jeff Geerling blog post from November 2024 on the power efficiency of Apple's M4 Mac mini, written amid his Pi 5 GPU and AmpereOne server testing.
Apple's M4 Max chip is the fastest single-core performer in consumer computing(twitter.com)
341 points by retskrad | 451 comments
An X (Twitter) post claiming Apple's M4 Max chip has the fastest single-core performance among consumer computers; the post content did not load.
SQLite: 35% Faster Than the Filesystem(sqlite.org)
374 points by yla92 | 127 comments
SQLite documentation page explaining that reading and writing small blobs in SQLite is about 35% faster and uses less disk space than individual files.
Vision language models are blind(vlmsareblind.github.io)
425 points by taesiri | 172 comments
Project page for the research paper "Vision language models are blind" from Auburn University and University of Alberta, with links to the arXiv paper, code, and dataset.
Getting 50% (SoTA) on Arc-AGI with GPT-4o(redwoodresearch.substack.com)
359 points by tomduncalf | 195 comments
Redwood Research blog post by Ryan Greenblatt describing reaching 50% on the ARC-AGI public test set by sampling thousands of GPT-4o-generated Python programs.
Cray-1 vs Raspberry Pi(roylongbottom.org.uk)
319 points by stare_spb | 164 comments
Roy Longbottom's page comparing Cray-1 supercomputer benchmark performance with home computers, phones, tablets and Raspberry Pi, plus his background in benchmarking.
Ripgrep is faster than grep, ag, Git grep, ucg, pt, sift (2016)(blog.burntsushi.net)
344 points by subset | 171 comments
Andrew Gallant's 2016 blog post introducing ripgrep, a Rust command-line search tool, with 25 benchmarks against grep, ag, git grep, ucg, pt and sift.
Fast self-hostable open-source workflow engine(windmill.dev)
346 points by rubenfiszel | 165 comments
Windmill launch week blog post claiming its self-hostable open-source workflow engine is the fastest, with benchmarks against other engines.
Ubuntu squeezes more performance than Windows 11 on new AMD Zen 4 Threadripper(phoronix.com)
319 points by marcodiego | 328 comments
Phoronix review benchmarking Ubuntu 23.10 against Windows 11 on an HP Z6 G5 A workstation with the 96-core AMD Threadripper PRO 7995WX, finding ~20% better Linux performance.
Asking 60 LLMs a set of 20 questions(benchmarks.llmonitor.com)
653 points by vincelt | 309 comments
A page showing raw answers from about 60 LLMs to 20 prompts testing reasoning, instruction following, and creativity, browsable by prompt or model.
Firefox has surpassed Chrome on Speedometer(treeherder.mozilla.org)
1077 points by akyuu | 462 comments
Mozilla Perfherder graph comparing Firefox and Chrome Speedometer benchmark scores on Windows 10 over the past year, showing Firefox catching up to Chrome.
Vale's first prototype for immutable region borrowing(verdagon.dev)
365 points by modernerd | 122 comments
Blog post by Evan Ovadia presenting results and benchmarks from Vale's first prototype of immutable region borrowing for reducing memory safety overhead.