Marko Simic

Building 🧠 for 🤖 ▶️ 🪙 👥

[Marko Simic @simicvm]

@chetaslua @TheRealAdamG gpt pro extended thinking, 32min run, same prompt

[Marko Simic @simicvm]

@MLStreetTalk M4 Max 128gb
llama-server -hf ggml-org/gemma-4-26B-A4B-it-GGUF:Q8_0
read: ~1000t/s
write: ~75t/s

[Marko Simic @simicvm]

@ivanfioravanti @danieltvela this command works for spec decoding but running it I get 2x lower t/s than when just running the big model alone 🤔 llama-server -hf ggml-org/gemma-4-26B-A4B-it-GGUF:Q8_0 -hfd ggml-org/gemma-4-E2B-it-GGUF:Q8_0 --no-mmproj

[Marko Simic @simicvm]

@ivanfioravanti opencode mac app is quite good. has a git worktree support out of the box

[Marko Simic @simicvm]

my qwen3.5-27b is not well... @lmstudio

[AGI House SF @AGIHouseSF]

Meet @svegas18, @AustinBaggio , @christinetyip from autoresearch@home:
> Wednesday, they release the most comprehensive autoresearch repo & leaderboard
> Same day, we plan and launch an autoresearch hack
> Thursday, they decide to fly y from Canada & NYC to present their work

[Marko Simic @simicvm]

here, made it for you @Kantrowitz @ranjanxroy

[Marko Simic @simicvm]

@edouard_iosdev @levelsio I've traveled to the US with it many times. I was only stopped once for a check. They took a sample, ran a test, and let me go within a few minutes.

[Marko Simic @simicvm]

@sawyerhood yeap, did that few weeks ago. runs qwen3 stt, all local
https://t.co/XmuCFsr3jN

[Marko Simic @simicvm]

@RobertTLange very cool work! it's great when it works, but so far in my testing I am getting mixed results on doc-to-lora.

[Marko Simic @simicvm]

We vibe-coded https://t.co/3dM07J4YC7 to test different OCR models. Between DeepSeek OCR 2, GLM OCR, and LightOnOCR-2, we are getting the most consistent results from DeepSeek. Its bounding box placement is also quite good.

[Marko Simic @simicvm]

@clattner_llvm @AnthropicAI Cool article! I'm curious, though, why is this the most interesting lesson if it’s exactly what you’d expect from an LLM? You even mention that in the next paragraph! 🙂

[Marko Simic @simicvm]

v0.2.0 now with the hotkey trigger remap functionality
https://t.co/W5WNbGSD6x

[Marko Simic @simicvm]

fully local Wispr alternative app running Qwen3 ASR 0.6B via mlx-audio-swift. practically instantaneous result in any active input field. @Prince_Canuma

[Marko Simic @simicvm]

@Prince_Canuma here is the link @MitchForest
https://t.co/XmuCFsrB9l

[Marko Simic @simicvm]

fully local Wispr alternative app running Qwen3 ASR 0.6B via mlx-audio-swift. practically instantaneous result in any active input field. @Prince_Canuma

[Marko Simic @simicvm]

opus 4.6 really likes its "belt-and-suspenders" phrase

[Marko Simic @simicvm]

@EuroJohannes @Prince_Canuma I like pageless layout and it looks better in gray

[Marko Simic @simicvm]

@Prince_Canuma running the same image on their demo app provides a good output. although it takes quite long to get the result, but that might be due to some free demo constraints.

[Marko Simic @simicvm]

@Prince_Canuma for example, I was not able to make it recognize all the text in this image, regardless of the mode I used. meanwhile, DS does it with ease.

[Marko Simic @simicvm]

@Prince_Canuma not sure how they got those high benchmark results 🤔

[Marko Simic @simicvm]

@Prince_Canuma did some tests with it. running on mbp m4 max 128. definitely faster than DeepSeek OCR2. however, it's significantly worse than DS on complex documents.