Max Idahl

This account is verified. Learn more

@maxidahl

AI Research @ ellamind

Joined January 2017

919 Following 195 Followers 114 Posts

Latest posts 5

Max Idahl @maxidahl 15h

prime has become my goto provider on OR for good reasons

Prime Intellect @PrimeIntellect 4d

Prime Inference has the highest cache rate of any third-party provider for DeepSeek V4.1 Flash 1017

High cache hit rate means that more of your prompt is billed at the cache-read price

This offers the lowest effective input cost on the board: $0.00443/M

View quoted post on X

Max Idahl @maxidahl 21h

im curious what made TML deviate from OpenAI's o200k tokenizer for Inkling 2
- same vocab size and pre-tokenizer regex
- new vocab, ~40% overlap with o200k
- less space+word tokens
- zero tokens that mix punctuation and letters (no more "'s" or "_loss" tokens)
- "\n" never shares a token with a letter, digit or punctuation

Thinking Machines Verified account @thinkymachines 1d

Inkling 2 Small’s weights are now available for download on HuggingFace under an Apache 2.0 license.

Make Inkling your own!
huggingface.co/thinkingmachin…
View quoted post on X

Max Idahl @maxidahl Oct 28

so many providers on @OpenRouter disqualifying themselves with ridiculously low output token cap. waky time, models now reason >16k tokens all the time.

Max Idahl @maxidahl Sep 16

Tokenization Tax Report 2026 now also on @huggingface
hf.co/spaces/ellamin…

huggingface.co Tokenization Tax Report 2026 - a Hugging Face Space by ellamind Use the interactive page to compare how many tokens different language models need for the same text. Pick a language, document type, or model from the navigation menu and view charts and tables th...

Max Idahl Verified account @maxidahl Sep 8

Compared the tokenizers of current models.

Efficient: Command A+, Muse (Glimmer and Spark, based on Llama 4’s vocab), Grok 4.5+, Qwen 3.5+, OpenAI o200k (GPT-6, GPT 5, gpt-oss)

Claude? +59% tokens for English text...

View quoted post on X

Max Idahl @maxidahl Sep 9

Forgot to add: As European AI tradition demands, none of the four European tokenizers we tested beats the o200k baseline averaged across the seven EU languages we experimented with.

Max Idahl Verified account @maxidahl Sep 8

Compared the tokenizers of current models.

Efficient: Command A+, Muse (Glimmer and Spark, based on Llama 4’s vocab), Grok 4.5+, Qwen 3.5+, OpenAI o200k (GPT-6, GPT 5, gpt-oss)

Claude? +59% tokens for English text...

View quoted post on X