We read it, so you don't have to.
Close readings of the documents, papers, and filings shaping the future, with the receipts.
ResearchThe two Gemma 4 31B entries: how Reki and forge play ARC-AGI-3
Second and third in the ARC-AGI-3 milestone both ran Gemma 4 31B on vLLM: four frames at eight times scale,…
InsightsInfinite play by eviction: how a 27B plays past its context window
Infinite play by eviction: how the winning ARC-AGI-3 harness keeps a 27B model playing past a 64K context by dropping…
ResearchThe Duck harness, read line by line: how a 27B won an ARC-AGI-3 milestone
The Duck harness won the first ARC-AGI-3 milestone with a 27B model and a Python REPL, scoring 1.21% official and…
InsightsTwo agents on one LM Studio endpoint: what concurrent requests cost
Two identical LM Studio concurrent requests to a 31B on a 48 GB Mac: each ran at half speed and…
AnalysisA REPL agent in its smallest form: one model call writes the player
A REPL agent in its smallest form: one local model call wrote a 44 line policy that played 60,000 ARC-AGI-3…
ResearchImage vs text for a local vision model: one board, two costs
Image vs text for a local vision model, measured: the same 64x64 board cost 4,179 prompt tokens as hex and…
InsightsA local model agent on one ARC-AGI-3 level: 40 moves, 230 minutes, nothing
A local model agent played 40 moves of one ARC-AGI-3 level in 230 minutes, 154,673 reasoning tokens, and cleared nothing.…
ResearchARC-AGI-3 on a Mac: what a 48 GB M4 Pro can and cannot do
ARC-AGI-3 on a Mac, measured on a 48 GB M4 Pro: the toolkit does 4,000 actions a second, a 31B…
ResearchARC-AGI-3 scores: under 1% is six months out of date
ARC-AGI-3 scores from 0.51% at launch to 62.7% shared harness and 99.9% vendor harness in September 2026, in date order…
AnalysisARC-AGI-2 vs ARC-AGI-3: what changed, and why the numbers do not transfer
ARC-AGI-2 vs ARC-AGI-3: static puzzles scored on accuracy against interactive games scored on action efficiency. Why a 24% and a…
InsightsReasoning tokens are output tokens, and five APIs count them differently
Reasoning tokens are output tokens on LM Studio, OpenAI, Anthropic, Gemini and OpenRouter, and each names and caps them differently.…
AnalysisLM Studio reasoning effort, temperature and schemas: what changes the thinking
LM Studio reasoning effort, temperature, JSON schemas and tool calls measured on one board: low effort thought most, a schema…
ResearchThe ARC-AGI-3 human baseline: all 183 levels, read from disk
The ARC-AGI-3 human baseline is the upper median first-time player, not the best. All 183 baselines read from disk: 6…
InsightsA random agent on all 25 ARC-AGI-3 games: 12 clears, no score
An ARC-AGI-3 random agent ran 60,000 actions on all 25 public games: twelve first levels cleared, two second levels, about…
AnalysisAll 25 public ARC-AGI-3 games compared, from source
All 25 public ARC-AGI-3 games downloaded and compared from source: lines of code, levels, action sets, human baselines, and what…
AnalysisHow an ARC-AGI-3 game runs: ARCBaseGame from source
ARCBaseGame is the 593 line class every ARC-AGI-3 game inherits. The action loop, why RESET means two things, and the…
AnalysisWhat is inside an ARC-AGI-3 game file
We opened an ARC-AGI-3 game file: 2,060 lines of runnable Python with 26 of 34 functions renamed to random letters.…
ResearchARC-AGI-3 is not saturated: which number to believe
Is ARC-AGI-3 saturated? Three numbers are circulating as a solved benchmark, and only one compares systems. Which set, which harness,…
ResearchRHAE explained: how ARC-AGI-3 scores an agent
RHAE is not accuracy. It scores your agent on moves against a first-time human, squared. Here is the formula, the…
InsightsARC-AGI-3 local model test: what one move costs
We ran the MIT licensed ARC-AGI-3 toolkit on a 48 GB M4 Pro with a local model, measured what a…
AnalysisAn agent’s system prompt tokens can outgrow the context window it runs in, by nearly five times
We counted the agent system prompt tokens in four real context files. Together they are 38,922 tokens, which is 475…
AnalysisRun two models at once on 48 GB and they both fit, but the thing that breaks is the one nobody measures
We run two models at once on a 48 GB Mac. RAM was never the limit: both fit in 26.78…
AnalysisA 64K context window costs 4.4 GB of RAM on top of the weights, and your model manager never shows it
Measured on a 48 GB Mac: context window ram usage for a local llm runs 80 KB per token. LM…
AnalysisPrompt processing speed on an M4 Pro is 65 tokens a second, and your benchmark is timing the cache instead
Measured on an M4 Pro: local llm prompt processing speed is 65 tokens a second. Repeat the same prompt and…
AnalysisTwo Unsloth UD quant names, one GGUF file, and not a single IQ1 or IQ2 tensor inside
An Unsloth UD quant name describes the recipe, not the GGUF tensors. Two Nemotron builds are byte-identical under different names,…
AnalysisNemotron 3.5 Lightning’s 4-bit GGUF is 12 percent 16-bit, and the weight is a head most people never switch on
The Nemotron 3.5 Lightning GGUF from NVIDIA is 22.46 GB, and 2.67 GB of it is BF16. That block is…
ResearchYour 4-bit vision model sees in 16-bit, and the best anyone offers is 8
An mmproj GGUF holds a whole vision encoder. Four publishers ship it unquantised at BF16 or F32, and the one…
ResearchThree publishers, one schedule: the quantisation policy we called a choice is a line of llama.cpp
A llama.cpp quantisation rule decides which layers get more bits. One line predicts all 20 promoted blocks in three publishers'…
ResearchThe speculative decoding head ships two different ways, and neither one is in the model you downloaded
An MTP draft model GGUF ships either as an extra block inside the model or as a separate 49-tensor file.…
ResearchExperts really do go missing from MoE models, and the builds that remove them say so in the metadata
REAP pruned MoE builds remove a quarter of the experts and declare it. We read four GGUF tensor tables: 48…
AnalysisTwo GGUF builds of the same model differ by a whole block, and it is the speculative decoding head
One Qwen3.5 GGUF build ships 41 blocks and another ships 40. The extra block is the MTP head for speculative…
AnalysisHugging Face ships a telemetry function nothing calls, and a header that reports your PyTorch version
Hugging Face telemetry has an opt-out env var and a send function the library never calls. The data trail is…
AnalysisWhich local AI tools give your machine a permanent name, and which only look like they do
A local AI machine ID turns anonymous requests into a profile. We swept five installed tools: a naive search says…
AnalysisEvery local AI app ships a crash reporter, and none of the three we checked turns it on
A crash reporter in a local AI app looks alarming in Activity Monitor. We checked three, and the crashpad process…
AnalysisOpenCode contacts Sentry before you type anything, and two other findings that were not real
OpenCode telemetry, measured on a live launch. A production Sentry DSN is baked into the app and it connects to…
ResearchThe smallest Kimi K3 you can download is 589 GB, and two builds of it disagree by 58
Running Kimi K3 local is not close to possible on a Mac. We summed every shard: the smallest complete build…
AnalysisLM Studio does not track you, and it routes every model search through its own servers
LM Studio privacy, tested on a live install. No telemetry keys and no outbound connections at rest, but model search…
ResearchOllama’s MLX engine has no 32GB requirement: the rule is three lines of Go
The Ollama MLX 32GB requirement is not in the source at all. Routing is decided by model format, not by…
ResearchEight GGUF builds of one model: the filename does not tell you what is inside
GGUF quant naming, audited across eight builds of one model. A file called UD-Q4_K_XL holds no Q4_K, and none had…
AnalysisBonsai 27B in a coding agent: a 1-bit model that calls tools cleanly and misses the line number
We put Bonsai 27B, a 1-bit model in 3.8 GB, behind the Pi coding agent. Tool calls came back clean…
AnalysisMoE on a Mac: 4 to 5 times faster, and bigger in memory than the dense model it beats
A 26B MoE decodes 4 to 5 times faster than a dense model on a Mac and can still take…
ResearchNVFP4 vs MXFP4: the same 4 bits, 36% apart
MXFP4 vs NVFP4 comes down to two choices: block size and scale format. Same 4-bit element, and NVIDIA reports MXFP4…
AnalysisNVFP4 vs Q4_K_M: the vendor published no numbers, and the independent ones favour Q4_K_M
NVFP4 vs Q4_K_M: the vendor published no numbers to check, and the one independent comparison has Q4_K_M ahead on KL…
AnalysisDoes llama.cpp support MLX? No, and a maintainer explained why in 2023
Does llama.cpp support MLX? No, and a maintainer said why in 2023. What Apple Silicon actually runs on, what Ollama…
AnalysisNVFP4 on Apple Silicon: an NVIDIA format that only runs on a Mac
NVFP4 on Apple Silicon is an NVIDIA format Ollama will only serve to macOS. Why the 412 error exists, what…
ResearchNVFP4 does not range from -8 to 7.5: the 16 values, corrected
The NVFP4 range is not -8 to 7.5. Its 16 bit patterns hold just 15 values, from -6 to 6,…
AnalysisThe largest distillation attack yet: the Alibaba numbers
Somewhere between late April and early June, a machine somewhere opened a fresh conversation with Claude roughly every seventh of…
ResearchBonsai 27B vs Gemma 4 31B on a Mac: which local model to run
Two models sat loaded in the same copy of LM Studio this week, and one of them was a fifth…
AnalysisBonsai 27B 1-bit on a Mac: a 27B model in 3.8 GB
Most 27-billion-parameter models arrive as a download you plan an evening around. The Bonsai 27B 1-bit build from PrismML lands…
AnalysisHow AI companies hide tracking code, from Claude Code to your local tools
AI tracking code hides in plain sight. We break down the covert tracker Anthropic shipped in Claude Code, the telemetry…
InsightsGraphify has 81,000 stars. We read the code instead of running it
Graphify has 81,000 GitHub stars in three months. We read the Graphify source instead of running it: zero telemetry, local-first…
InsightsGraphify on a local model: we ran it and watched the network
We ran Graphify on a local model and watched every socket it opened. 358 went to LM Studio, none went…
AnalysisOpenCode vs OpenJarvis: which local AI agent phones home?
OpenCode vs OpenJarvis: we captured what each one sends home. One has a working off switch. The other, the agent…
ResearchOpenJarvis on a Mac: we retested Stanford’s local AI agent
We reinstalled OpenJarvis on a Mac, captured its telemetry with a local sink, and found the undocumented off switch. What…
InsightsBrand voice for AI agents: the layer that beats a better model
We built a five-file brand voice for AI agents. Why the output spec, not the model, is the layer that…
ResearchHow exo Clusters Macs to Run a 671B Model Locally
exo is open-source software that pools several Apple Silicon Macs into one machine, so a cluster can run models like…
ResearchSet up local LLM tool calling in Hermes and OpenCode on a Mac
We set up local LLM tool calling in Hermes and OpenCode on a 48 GB Mac mini. Why small models…
ResearchM4 Pro vs M5 Max for local inference on Apple silicon
M4 Pro vs M5 Max for local inference. Memory bandwidth, unified memory ceilings and time to first token, and whether…
AnalysisLocal vs cloud AI: what to run where in 2026
Local vs cloud AI in 2026. A decision guide for what to run on your own hardware, what to send…
ResearchGemma 4 E4B on edge hardware: small models catch up
Gemma 4 E4B runs on 8 GB of unified memory. What a small model can now do on edge hardware,…
AnalysisWhat Ollama NVFP4 means for local model quality
Ollama NVFP4 and q4_K_M are nearly the same size on disk. What the newer 4-bit format changes for local model…
InsightsGemma 4 26B MoE local: quality per gigabyte on unified memory
Gemma 4 26B MoE runs from 16 GB of weights. We look at quality per gigabyte on unified memory, and…
ResearchGemma 4 31B vs Qwen 3.5 27B on unified memory
Gemma 4 31B vs Qwen 3.5 27B on unified memory. Which fits, which is faster, and which answers better when…
AnalysisMLX vs llama.cpp vs MetalRT: best inference engine on Apple Silicon in 2026
MLX vs llama.cpp vs MetalRT: one quantized model, three engines, one Mac. Generation speed, memory and setup, measured rather than…
InsightsLM Studio vs Ollama on Mac: which one to use in 2026
LM Studio vs Ollama on a Mac in 2026. Setup, speed, model management and the API each one exposes, plus…
ResearchLocal LLM benchmarks on M4 Pro: Gemma, Qwen, Llama speeds
Local LLM benchmarks on an M4 Pro. Gemma, Qwen and Llama measured on tokens per second, memory used, and the…
InsightsWhat the Ollama MLX shift means for local AI on Mac
Ollama is moving toward MLX on Apple Silicon. What the shift changes for speed, memory and model choice, and whether…
AnalysisBest Local LLMs for Coding on a Mac in 2026: Benchmarked and Ranked
We benchmarked and ranked the best local LLMs for coding on a Mac in 2026, on real refactors, with speed,…
InsightsPrivacy-first coding: why you should run AI agents entirely offline
Your config files hold keys and endpoints. Here is why we run AI agents offline, what a cloud agent can…
ResearchHow to choose between Ollama, LM Studio, and MLX for local models
Three ways to serve a local model on a Mac. We compare Ollama, LM Studio and MLX on setup, speed…
AnalysisThe Hidden Costs of ‘Free’ Cloud AI Tiers
A free cloud AI tier does not cost zero. It costs something other than dollars: your data, your reliability, your…
InsightsAgentic AI explained: what it changed for local AI
Agentic AI explained without the hype. Stripped down it is one real shift: a model that answers questions became a…
InsightsLocal Deep Research vs Perplexity and the Cloud
Local deep research vs Perplexity: a research engine on your own machine keeps every query private and charges nothing per…
InsightsOpen WebUI vs Jan: The Local ChatGPT, Picked
Open WebUI vs Jan: two ways to put a private ChatGPT in front of someone. A feature-rich browser workspace, or…
AnalysisDo you need an agent framework? smolagents, and when to skip it
Do you need an agent framework at all? smolagents vs the alternatives, and why most of the noise is selling…
InsightsIs Local AI Good Enough to Replace the Paid Tools?
Is local AI good enough to replace the paid tools? For most of what you do in a day, yes.…
ResearchWhich Local LLM Should You Actually Run in 2026?
The best local LLM in 2026 is not the biggest one. It is the model that fits your RAM with…
AnalysisCloud AI vs local: the true cost compared
Stop arguing about whether local AI is cheaper. Add up the real cost of AI subscriptions vs local: a machine…
ResearchOllama vs MLX vs Jan: Running Local Models on a Mac
Three ways to run a model on Apple Silicon: the foundation, the fast one, and the friendly one. We run…
AnalysisAider vs Cline vs OpenCode vs OpenHands, Picked for You
Four open-source coding agents, four different answers to one question: where do you want to work? Terminal, editor, app window,…
AnalysisOpenCode vs Pi vs Cursor: Which Coding Agent Do You Need?
Three coding agents, three different bets on where you work and where your code goes. We have run all three.…
AnalysisBest local model for OpenCode, compared
Nemotron vs DeepSeek for local coding. DeepSeek scores higher on pure coding benchmarks, and we still run Nemotron as the…