It is great to see another player introduce a fully open stack. Nvidia's Nemotron is the only other prominent one I know of.
All that said, the headline claims do not match the self-reported performance. For example, the dense 32B model is significantly behind Qwen3.8 27B (chart towards the bottom of https://ifm.ai/blog/k2). Gemma4 31B is not in the comparison set. This is the most important sweet spot for self hosted open-weight models today and real competition here will be very welcome.
They have the 32B listed as "stage 1" with the note "final checkpoint to be released." So, not finished yet. Not sure why you'd release it if it's not finished, but that's the explanation.
Looks like the strategy of regular updates with incremental improvements is working out well. Interestingly, the biggest jump in Artificial Analysis Intelligence Index score is for reasoning level Medium ( 3.7 was 51, 53, 57 for Low, Medium and High, 3.8 is 52,57, 59 respectively). I think scores at lower reasoning levels are more indicative of model capability since higher reasoning levels are focussed on benchmaxxing. We use the lowest reasoning level in production with good results.
I'm not an expert but I agree with your statement on the lower reasoning levels.
Lots of models seem to just allow the model to "bloatmax" tokens in order to get bumps at high/max reasoning levels. Many of the max reasoning levels allow models to use up to double or more the tokens the next lowest reasoning level uses. Its basically only useful for people who have no cost or time stipulations on anything.
I think I actually preferred it when we had models that either had reasoning enabled or didn't.
Quoting RGFusion from Reddit: LLMs run into an issue where the further you train a model, the more it overwrites facts with generalized concepts. You need the model to be able to do both. Intelligence arises from generalization, but without accurate information the model will hallucinate.
The engram table allows for a low-computational method of fact-recall. You can think of it like a better form of RAG, where the data doesn't take up any of your context window and it's injected deeper into the model's layers, freeing the lower layers to carry out abstraction. This results in better "focus" for the model, both in regards to its intelligence and context recall.
Basically, they've separated the specificity-critical portions of the models memory into a parameter space that doesn't need fast compute (you can run it on system RAM) and allows the model to be trained on higher volumes of data without ruining its knowledge-base.
There is a difference between needing frontier capability because one is solving a truly open ended problem, and needing a reliable workhorse model to do something well understood. Local models (like Qwen 3.8 27B) have gotten so good that they can do all routine tasks at a fraction of the cost of frontier models.
This is open model intended for local usage. So, it doesn't matter. If the end user wants to run this on their own hardware, cost is defined by the electricity price + the price of the processing power. Providers can increase the price as long as the users don't switch for buying the hardware themselves instead.
Even a 4-bit quant of Qwen3.8 27b is indistinguishable from Gemini 3.7 flash in our internal tests. With an RTX5090 card and ninfer, you can get ~800 TPS token generation (c=8) and ~140 Tokens per second single stream.
Amit Patel was employee #7 at Google and built many things including the little geo display in the lobby that would show live search queries as dots color coded by language.
And writing “Don’t be evil” everywhere. I suspect being the “originator” is contested; but this wonderful nerd, who made Red Blob Games, is the spirit.
Barren Realms Elite and Falcons Eye were my favourite doors games as a teen (written by his brother). Inter-BBS - playing against people in other countries - was just amazing at the time.
Interesting. I found this article that does point to Amit Patel as the originator of “Don’t be evil” at google.
“Mayer explained that Don't Be Evil was coined in 1999 by one of Google's first engineers, Amit Patel, who shared a work cubicle with Mayer.
She said Patel and other early employees were resistant when staff with business skills began joining the engineer-driven company. The engineers feared they would be pressured into moving certain clients higher in search results listings or building products they did not want to build.
Patel voiced his fears via the whiteboard in the conference room where Google sales people met clients.
"In this incredibly neat handwriting in tiny little letters on the bottom right hand side of the white board, he wrote 'Don't Be Evil'," Mayer said.
Some time later, when Google had 200-300 employees, its human resources team decided the company needed corporate values. They called a team meeting of some of Google's senior staff.
After they brainstormed 10 positive corporate values, Paul Bucheit, the inventor of Gmail, suggested ditching those in favour of one that covered all angles - Don't Be Evil.
"Originally Don't Be Evil was No.6 on the corporate values [list] but it's the only one that stuck because it's the catchiest and, as Paul said, encompasses everything else," said Mayer.”
I was lucky enough to spend ten or so hours with Amit over the span of six or eight years after he left Google. He has a fantastic ability to find the simple core of very complicated things, making it easy for others to understand. People repeatedly thought it was so impressive that I understood deep concepts so thoroughly. Then I'd say, "wait, let me draw the simple line chart on the whiteboard that Amit drew for me." Then they immediately understood it just as well as I did. My time with him always left me feeling inspired. He's a truly nice guy too.
The pricing is 40% of Kimi K3 at the same Model size. It will be interesting to see benchmarks tomorrow and to see how the pricing evolves across this size class.
Seminanlysis is estimating sub $1 cost per MT for ~2Trillion models. The numbers change based on throughput and quant, but it is conceivable that provider costs at scale are low enough that even $2.42 per MT on GLM 5.2 (current best price) is margin positive by a wide margin.
This resonates with my experience. At Morph we use gemini for well specified point coding tasks, and it does very well across millions of lines of code every day. We also use claude code as an engineering tool for our own codebase and it does better at being adaptive and for working on open ended issues.
All that said, the headline claims do not match the self-reported performance. For example, the dense 32B model is significantly behind Qwen3.8 27B (chart towards the bottom of https://ifm.ai/blog/k2). Gemma4 31B is not in the comparison set. This is the most important sweet spot for self hosted open-weight models today and real competition here will be very welcome.
reply