Sid Dev

Chewing on life.

[Sid Dev @sid_dev1]

Really interested to see how this ~74b MoE lands once RL is completed, as it seems like ~30b is the current local hotspot. Moreover, I'd presume they are going to convert to TiDAR for specdec speedup. Could be interesting!

[Sid Dev @sid_dev1]

@witcheer cool! Did you play around with the temp you ran the model at?

[Sid Dev @sid_dev1]

Kinda funny "The supported languages include: English. Improved using Qwen."

[Sid Dev @sid_dev1]

@DevaBuilds @Sapient_Int Naive goal but was trying to see if I could get the base model + HRM architecture to push into to Bonsai / LFM / Gemma territory in terms of performance (based on the simple benchmarks I did on them). I wasn't eval'ing on latency, purely curious about accuracy / answer quality.

[Sid Dev @sid_dev1]

I messed around with fine tuning @Sapient_Int 's HRM 1B model. I find HRM fascinating! Goal was to learn how fine tuning works and see if I could modestly improve off the base model.

Note the benchmarking is crude as I'm basing it off my workflows and severely hardware limited.

[Sid Dev @sid_dev1]

@0xSero Think people are sleeping on @ZyphraAI and smaller Nvidia models (self speculation / diffusion etc) wrt Deepseek-esq envelope pushing

[Sid Dev @sid_dev1]

Honestly I like this minimax m3 / opencode combo much better than antigravity / flash 3.5, probably not controversial

[Sid Dev @sid_dev1]

Really impressed with @opencode / @MiniMax_AI M3 combo. Used it for the first time and it pulled Nvidia Diffusion 3B 4 bit MLX out of a ditched by writing a custom MLX loader, ran it on my tiny M2 mini 8gb, and came up with some great insights.

[Sid Dev @sid_dev1]

Well this shows how bad Stepfun 2.7 and other stuff I've been using have been given how other people feel about this combo https://t.co/dkrIGg7n61

[Markets & Mayhem @Mayhem4Markets]

The market's giving MiniMax-M3 a vote of no confidence, as shares tumble down a massive 15.71% after running up into the M3 release.

Hearing from more and more people that this feels like a benchmaxxed model.

Again, hope to be proven wrong. But this seems quite bad.

[Sid Dev @sid_dev1]

@PavloMolchanov My apologies! I meant does it *only* apply to smaller deployments w/ less concurrency (and more free cores). But I think you answered both!

[Sid Dev @sid_dev1]

@itsPaulAi Had antigravity send it

[Sid Dev @sid_dev1]

@itsPaulAi Here's what I got so far

[Sid Dev @sid_dev1]

@itsPaulAi E4B QAT still crashed it without limiting GPU offload, etc. But the E4B lite version cruised at ~27tps!

[Sid Dev @sid_dev1]

@PaulGugAI @Teknium @NeoAIForecast M2 mini 8gb checking in.

[Sid Dev @sid_dev1]

@osanseviero This is awesome, thanks for thinking of the little guy / edge.

[Sid Dev @sid_dev1]

Here's the AI written summary of the fine turning journey (a bit too tired to do this by hand right now).

https://t.co/lnrkrBzTGb

[Sid Dev @sid_dev1]

Update, swapped to Deepseek v4 flash...it crushed it and then some.

[Sid Dev @sid_dev1]

@OnlyTerp Currently trying to have Nemo 3 Ultra (Hermes harness) simply update a blog post, write a new blog post, and upload to my website and it is looping badly with a failed tool calls. DSv4 flash / Mimo 2.7 etc handled this repeatable task no problem.

[Sid Dev @sid_dev1]

@Teknium Loving the desktop app, kudos to the whole team.

[Sid Dev @sid_dev1]

@OnlyTerp Currently trying to have Nemo 3 Ultra (Hermes harness) simply update a blog post, write a new blog post, and upload to my website and it is looping badly with a failed tool calls. DSv4 flash / Mimo 2.7 etc handled this repeatable task no problem.