Christian Merrill

@kalomaze But it’s for safety. Imagine how unsafe it would be to allow you to map text to tokens, we couldn’t have that.

hey isn't it kind of messed up that API customers of Opus 4.7 and Opus 4.8 are paying ~1.41x as much for general english output (when measured against a consistent tokenizer baseline) vs Opus 4.6? from THE ONLY lab that's cagey about releasing the tokenizer... huh. that's... odd

@keennay I’m in the process….

welp I think I'm good to go with open-weight models if the plug ever gets pulled

@MiaAI_lab @QuixiAI @deepseek_ai @StepFun_ai I’m not sure I’ve benchmarked it well enough to tell you but the average latency to a Hermes agent is about 35sec per my dashboard. (I keep a proxy sidecar that counts all tokens and aggregates on my Mac mini) I should configure it to track more thou.

@M_Chimiste @QuixiAI @deepseek_ai @StepFun_ai Insane. How much tok/s when in deep context?

DeepSeek-v4-Flash beats Step-3.7-Flash in head-to-head tool calling benchmark. Full results in: github.com/MiaAI-Lab/Deep…

@MiaAI_lab @QuixiAI @deepseek_ai @StepFun_ai This is how it’s currently configured with the weights being stored on dedicated M.2 drives on the side. I probably should change the configuration since I believe it’s slower with them stacked like this but it’s more convenient space wise.

@MiaAI_lab @QuixiAI @deepseek_ai @StepFun_ai or for smaller models like m2.7 I run two instances

@QuixiAI @MiaAI_lab @deepseek_ai @StepFun_ai If I had a better way to run K2.6 I probably would. Though even though it’s slower than an Nvidia farm, it’s the best I’ve got so I kinda need it still 😅

@M_Chimiste @MiaAI_lab @deepseek_ai @StepFun_ai Wanna sell me one? 😁

@xlr8harder Thank you for using the correct abbreviation.

making fable NOFORN is just part of trump's universal basic honeytrap program

@Sentdex Once they are, it'll obviously be added to my list

@Sentdex I don't think GLM 5.2 is available as weights right?

While closed source AI is in shambles, open source is having one of the best weeks of all time. Z ai GLM 5.2 Minimax M3 Kimi 2.7 code

@KeyTryer They absolutely could challenge it but it would also risk undermining their regulatory push. The most similar case I can think of is Bernstein V US on designating cryptography a munition. Though it doesn’t help they’ve been comparing their their to a munition…

There's shockingly little discussion about Anthropic's legal recourse about whether the order was illegal or unconstitutional. I don't know much about this, but my guess is that they'll challenge it as soon as possible, just like they challenged the supply chain risk designation.

@LLMSherpa Based on that shirt alone I felt compelled to ensure the aura of the shirt matched the aura of the background.

I want to start a charity called "Beards for Billionaires", because the beard really does smooth over a lot of these problem areas. This is what I look like when I shave my beard.

I feel bad for Dario. Not because of the Mythos drama, but, like... Just fucking look at him. If I asked a caricature artist to draw the biggest, most over-the-top dorky looking guy they could possibly imagine, it wouldn't come close to the reality of Dario. Bless him. 🫶

@ivanfioravanti As I think about it maybe if miniaturized might work well for local models on cell phones since they deprecate pretty fast in comparison to say a server.

@ivanfioravanti This is really cool, but I have no idea how well this will scale.

Amazing experiment by Fabio! 🔥

@TheAhmadOsman I bought 48Tb of storage yesterday….

Before AGI arrives: Acquire GPUs. Go into debt if you must. But whatever you do, secure the GPUs.

@MiaAI_lab @QuixiAI @deepseek_ai @StepFun_ai I somehow had the foresight to buy two 512gb Mac studios. Not realizing the world was about to enter a crazy hardware phase.

@M_Chimiste @QuixiAI @deepseek_ai @StepFun_ai I wish I had the compute to run MiniMax M3. For now, DeepSeek-v4-Flash is unbeatable for 2x DGX Spark setup.

@QuixiAI @MiaAI_lab @deepseek_ai @StepFun_ai I had a lot of tool call issues with Step 3.7. I think I was using Q8 at the time in Hermes Agent. I ended up reverting to Minimax M2.7 and working on moving to M3 for the multimodal input.

@MiaAI_lab @deepseek_ai Deepseek v4 Flash is text-only, 284B @StepFun_ai Step 3.7 Flash is a Text + Vision model, 198B The vision and the smaller size are more appealing. I choose Step 3.7 Flash.

@Sentdex @UnslothAI @MiniMax_AI My friend, I have tokens and time and I want to play with M3 as well. 🤣 I’m working on a bunch of things wrt harnesses and coding and science and this just seems to align.

@M_Chimiste @UnslothAI @MiniMax_AI I dont know if it'll work ootb wit anything but minimax m3 but shouldnt be too hard to allow for others. M3 has some special tokens stuff goin on that im deliberately handling for. Eventually ill make it function with others too maybe. Rly just wanted to play with m3 lol

After spending too many hours trying to implement fixes for MiniMax M3 native tool calls serving via llama.cpp to work in existing agents, I simply had M3 write its own mini coding agent I'm calling: Minion Now my minion just edits itself to give me what I want as a coding agent and it works surprisingly well. Lots of changes I plan to make, feel free to use it if you like...but mostly it has me questioning if we all should just make our own agents at this point. Maybe MiniMax is exceptionally good at tool calls out the gate to make this super simple, but I am enjoying making my minion exactly what I want and nothing more! and it doesn't take 50K context to say "hi" (yet) We'll see how long I can keep my own bloat at bay. Also MiniMax M3 overall so far has me very impressed. This is a VERY cool model!

If you have the NVMe Go download as many models as you think you might ever want. Now, go on Huggingface. They’re coming for open models next.

@Sentdex @UnslothAI @MiniMax_AI I’m going to fork and also add code2lora. Will report back.

Github: github.com/Sentdex/minion MiniMax M3 GGUF models (im using MiniMax-M3-UD-Q4_K_XL): huggingface.co/unsloth/MiniMa… Thanks to @UnslothAI for making these avail so fast and of course to @MiniMax_AI for this epic model and making it open weights!