Reasoning

Cerebras Deploys Trillion-Parameter AI Models

Cerebras fits Kimi K3 (3T) across 59 wafers in a pipeline, serving 1000-2000+ tokens per second

DeepInfer: High-performance inference framework for DeepSeek models

DeepSeek open-sourced DeepInfer, its production inference engine, under Apache-2.0.

Mercury 3: Diffusion at scale

Scaling diffusion LLM 40x gives Mercury 3: 4–8x faster than speed-tier models, similar quality

DeepSeek banned from U.S. federal civilian agencies under new DHS order

DHS order bans DeepSeek from U.S. federal civilian agency IT systems

Timestral: Mistral's time-series foundation models

A 209M time-series model ranks first on three benchmarks and runs on CPU.

DeepSeek V4.1 Flash 1017 (Max)

DeepSeek V4.1 Flash 1017 Max scores 45 on the Intelligence Index at 211 tok/s and $0.30 input.

Hybrid Reasoning Agents: LLM agents that think in latent space, act through text

Mixing latent reasoning with text tool use cut tokens by up to 54.7% and lifted performance 10.4%.

DeepSeek-v4.1-Flash-1017

DeepSeek V4.1 Flash minor update lifts coding and agent benchmarks while cutting run time and output length

Show HN: Llmspeed – feel how fast a local LLM runs before you buy the hardware

Pick hardware and a model to preview local LLM speed before buying

Llama.cpp on a jailbroken PS5: run any GGUF model natively on the console

A jailbroken PS5's GPU runs llama.cpp as a home LLM server

A Plausible Explanation Is Enough to Mislead Multimodal Judges

Adding plausible explanations to fixed answers cut MLLM judges' wrong-answer rejection by 12.7–13.8 points.

OpenGCUA: Toward a Fully Open Stack for Generalist Computer Use

Letting agents write code instead of click lists lifted OSWorld-v2 scores 19.6 points and cut steps 53%.

LFM2.5-24B-A2B: Built for fast on-device agents

A 24B MoE with 2B active params matches 35B-class models on 5 of 8 agentic benchmarks.

Prune the Summary, Not the History: Source-Judged Compaction for Long-Horizon Agents

Pruning source-less entries from compaction summaries raised Pass3 by 2.1–8.9 points

What agents mean for Rust (and vice versa)

Rust's designer argues agent coding calls for stronger types and proofs, not less language design.

Whose turn is it? Context-guided model routing for stateful agents

A router that reads session history passed 3.3 points more tasks than one without it.

Gemini 4 Argon

Gemini 4 Argon ties #1 on CWE-Bench (68%) and sped up Gemma 4 E2B on vLLM 2.85x

Redesigning source control at Anthropic

Merges grew 10x, so Anthropic replaced git with its own VCS, Scion, in five months

Scaling Inference-Time Compute with Teams of Coding Agents

Delegating only planning to workers matched full-delegation reward at half the cost.

Janas: Run LLMs larger than RAM by streaming experts from SSD

Streaming MoE experts from SSD ran an 80B model on a 32GB laptop at 24 tok/s versus llama.cpp's 2.8.

Rewriting Prime Agent in Rust

An agent swarm ported its own TypeScript to Rust, cutting time-to-input about 14x.

Can you use autoregressive diffusion to generate market data?

Flow matching and atom smoothing helped diffusion handle spiky market data distributions

Show HN: Jevman – AI decision models play Pac-Man

Open-source Pac-Man benchmark ranks decision models over 100 real-time games each

Sub-1-Bit LLM Compression via Latent Factorization

Factorizing weights into binarized low-rank factors compresses LLMs down to 0.1 bits per weight

Why isn't the industry freaking out about DeepSeek 4.1 Flash?

A month on DeepSeek 4.1 Flash felt frontier-grade at a far lower cost, says the author

Learn2Play Bench: How Well Do LLM Agents Learn from Experience in Unfamiliar Environments?

On 20 unfamiliar-rule games, keeping raw history (MEMORY) beat more elaborate summarizing self-evolving methods.

Claude Haiku 5.5

Haiku 5.5 costs about 75% less than Haiku 4.5 on average and runs faster.

Write Like It's 1866: LLMs Relearn Telegraphese

Telling LLMs to write in telegraphese cut output tokens 40–49% with comprehension intact

WebFovea: When the Model Is Right but the Click Is Wrong -- Reliable Round Trips for Vision-Based Web Agents on Live Websites

Fixing the harness, not the model, raised a web agent's challenge score from 31.0 to 57.0

OpenTPU – An open-source AI accelerator, developed by AI

An AI-designed open-source accelerator runs real LLMs like Qwen3 on an FPGA card

Mistral Large 4

Mistral Large 4: 1T-parameter multimodal model, API now, open weights end of October

Dust: Pretraining Transformers Without Backpropagation

Adding per-token noise to activations lets transformers pretrain without backpropagation, rivaling backprop.

Base Models Can Reason By Taking a Cue From Training Data

Prefilling just two opening tokens lifted Olmo-3-7B's MATH-500 accuracy from 42% to 78%.

Run Qwen 3.8 Flash Next (125B) on consumer hardware (RTX 4090) at 100T/s

A local engine runs a 125B model on a 32GB gaming PC with OpenAI/Anthropic-compatible APIs.

Aleph Alpha Kolibri: How the sovereign German LLM works

A 78B MoE using only 3.5B per token, built for German RAG and self-hosting.

From the creator of Redis; run LLM locally with ds4

antirez's C engine ds4 runs DeepSeek V4, Qwen and GLM locally on Mac and CUDA

DeReAct: Decomposed Reasoning and Acting for Reliable AI Agents

Splitting ReAct into action gating and completion checks lifted weaker models' Pass@1 by up to 7.0 points.

Clef: Open-weight decision models, and new RL fine-tuning platform

Cloudflare releases Jev-compatible decision models Clef and Clef-flash, claiming the top Jev Decision Index spot

Context Language Models

Letting models edit their own context file beat SOTA context management with less compute

Initialization Improves LLM-Driven Discovery

Seeding iterative LLM discovery with a parallel-explored population lifted SOTA harness scores 2–21%.

Gemini 4 Argon (High): Intelligence, Performance and Price Analysis

Gemini 4 Argon (High) scores 53 on the AA Intelligence Index but is slow, with a 48s time to first token

Launch HN: Magnitude (YC S25) – Self-optimizing inference engine for agents

Open-source engine tunes kernels per model on your hardware, claiming 2x Ollama speed on average

OpenDLSS: A Vulkan Reimplementation of Nvidia's DLSS 5 Neural Rendering Network

A Vulkan reimplementation of DLSS 5's neural rendering net that matches the original byte for byte

Show HN: TurboGPT: train 22KiB transformer in 13s

A CUDA C++ project trains a tiny byte-level GPT in under a minute.

ESP32S3 cluster running 1.58-bit (BitNet) Language model

A 7-node ESP32-S3 cluster runs a 0.4B BitNet LLM over an SPI daisy-chain

How Much Prompt Is Enough? A Blackbox Minimization of Few-Shots in LLMs

Delta debugging trimmed few-shot prompts by 65.3% on average with unchanged propositional output

Language Models Are "Insecure" Reporters

Adding "Be honest" raised GPT-5.5's flagging of a hidden negative result from 2/200 to 190/200.

MicroLLM Lab – Try 7 tiny LLM's in the browser

Run 25M–360M LLMs in-browser via WebGPU and benchmark their speed and accuracy

Prompting Claude Opus 5.5

Opus 5.5 at medium effort matches Opus 5 at high; guide covers prompt and harness tweaks

Faster prompt lookup drafting in llama.cpp

Removing map copies and swapping data structures made llama.cpp prompt lookup drafting up to 42x faster.

DeepSeek Elastic Compute (DSec)

DeepSeek's DSec runs about 3M sandboxes a day on 160 nodes for agentic training

A single function Jev-like wrapper for LLMs, including vision models

Forcing a one-letter answer and reading logprobs turned webcam frames into classifiers

Reasoning Concentrates Errors, and Self-Consistency Never Notices

Toggling only reasoning on the same weights made wrong answers collide, and voting couldn't filter them.

Thinking Less to Simulate Better: Intuitive Prompting Improves LLM Agents Simulating Individual Social Media Reactions, Including Unfamiliar Content

Telling LLM agents to answer instantly and intuitively predicted individual reactions best.

Best LLM for every budget, updated daily

A daily-updated site plots LLM benchmark scores against price to show the best model per budget.

Persuaded, Not Informed: Incentive-Misaligned Witnesses Defeat In-Context Grounding

LLMs believed optimistic sales reps; moving the math to code caught 28 of 31 bad deals.

When Should LLMs Trust Their Own Revisions? A Risk-Aware Study of Intrinsic Self-Correction

Asking 29 LLMs to revise answers fixed errors but also flipped correct answers

Do We Need Complex Topology Control? Distinct-Peer Random Routing Improves Cost-Efficiency in Sparse Multi-Agent Debate

Random distinct peers each round plus a 3-round cutoff cut debate tokens by about 78% with similar accuracy.

CliffCompaction: Cost-Efficient Compaction for Long-Horizon Coding Agents

Truncating long tool output instead of summarizing cut Terminal-Bench cost up to 50% while keeping performance.

Claude Opus 5.5 Intelligence, Performance and Price Analysis (Max)

Opus 5.5 max tops the Intelligence Index at 58 but is verbose and pricey.