News
Alibaba opens Qwen3.8-Max, without the eyes
The weights published on 9 August take text only at 262,144 tokens; the hosted version keeps vision input.
7 minAI
Everything we publish on AI: the models we run on our own machines, what the tools send home, the benchmarks we measure ourselves and the claims we check against the code. Start with the investigations, then the guides.
News
Claude Fable 5.1 costs the same, cache reads cost less
Released 1 September at $10 and $50 per million tokens, cache reads cut to $0.25, and three breaking changes for Fable 5 code.
7 min
Insights
How a 27B plays past its own context window
The winning ARC-AGI-3 harness drops old turns, keeps facts in Python and keeps going. Infinite play by eviction, read from the code.
6 min
Guides
A picture costs a local model 22 times less
The same 64 by 64 board was 4,179 prompt tokens as hex and 187 as a PNG, with less than half the reasoning.
12 min
Guides
What a 48 GB Mac can do on ARC-AGI-3
The toolkit runs 4,000 actions a second; a 31B reasons at 11 tokens a second; one move costs 67 to 443 seconds.
16 min
Insights
Five APIs count reasoning tokens five ways
LM Studio, OpenAI, Anthropic, Gemini and OpenRouter all bill them as output and name, place and cap them differently. One client reads all five.
7 min
Guides
A first ARC-AGI-3 agent in 40 lines
It survives resets, converts action ids, handles clicks and scores locally. Every guard was paid for by a failure first.
18 min
Insights
ARC-AGI-3 is not saturated
Three numbers circulate as a solved benchmark and only one compares systems. Which set, which harness, and what each cost.
17 min
Guides
Why LM Studio returns an empty reply
Reasoning ate 4,093 of 4,096 tokens, four times in a row. Here is the fix.
22 min
Guides
What one ARC-AGI-3 move costs a local model
The MIT-licensed toolkit on a 48 GB M4 Pro with a local model, a single move measured, and who is winning the prize.
12 min