Benjamin Marie

@bnjmn_marie

Independent AI researcher (LLM, NLP). My blog, The Kaitchup - AI on a Budget: https://kaitchup.substack.com/

following
followers

MiMo-V2.6 reportedly scales reinforcement learning to about 2 billion tokens per step

In a September 16 update, the team behind MiMo-V2.6 said it was scaling compute, agent environments and grading, with plans to open-source details in stages.

38 Sources, 27d ago, updated 27d ago

Tester favors Unsloth for speed among Qwen3.8-27B NVFP4 variants

In one user's long-context coding tests on an RTX Pro 6000, NVIDIA's version with MTP-4 was roughly 2.5–3× slower than Unsloth, despite being about 2× faster than running without MTP.

1 Source, 37d ago, first seen 37d ago

AI benchmarks as a measure of peak capability

A user calls V4.1 “very brittle,” claiming it can see “massive gains” when AGENTS.md instructions are optimized for a particular scenario.

41 Sources, 24d ago, updated 24d ago

The Next Web reports a 99.9% vs. 62.7% benchmark gap for the same model and test

The Next Web says OpenAI published the higher score, while the benchmark’s own software returned the lower one. The model and test were the same; the surrounding software setup differed.

3 Sources, 42d ago, first seen 42d ago

Benjamin Marie Questions GPQA Diamond Benchmark Usage by OpenAI

Independent researcher notes persistence of older evaluation method for advanced AI models.

1 Source, 47d ago, first seen 47d ago

@scaling01 States Fable Is Cooked

Pseudonymous AI commentator @scaling01 asserts the claim about the model in a post.

11 Sources, 47d ago, first seen 47d ago

Artificial Analysis Ranks Muse Spark 1.3 Most Cost Efficient

Artificial Analysis calls out its edge on cost per intelligence task.

38 Sources, 48d ago, first seen 48d ago

Featured in

MiMo-V2.6 reportedly scales reinforcement learning to about 2 billion tokens per step

Tester favors Unsloth for speed among Qwen3.8-27B NVFP4 variants

AI benchmarks as a measure of peak capability

The Next Web reports a 99.9% vs. 62.7% benchmark gap for the same model and test

Benjamin Marie Questions GPQA Diamond Benchmark Usage by OpenAI

@scaling01 States Fable Is Cooked

Artificial Analysis Ranks Muse Spark 1.3 Most Cost Efficient

Featured in

MiMo-V2.6 reportedly scales reinforcement learning to about 2 billion tokens per step

Tester favors Unsloth for speed among Qwen3.8-27B NVFP4 variants

AI benchmarks as a measure of peak capability

The Next Web reports a 99.9% vs. 62.7% benchmark gap for the same model and test

Benjamin Marie Questions GPQA Diamond Benchmark Usage by OpenAI

@scaling01 States Fable Is Cooked

Artificial Analysis Ranks Muse Spark 1.3 Most Cost Efficient