I have belief in OM in directional future. But can’t really test it without company enable it to really do side to side comparisons.
It’s also harder to do so because the agentic workflow already tailored so much to CC setup, I don’t even get time to Codex even. It’s not like no time to work on codex, but I can’t wait to keep iterating on CC as it’s not done yet.
Personal projects most of sonnet level is pretty easy because of the complexity and context limits.
The open-source angle is neat, but framing it as a global leveller misses what happens inside China’s firewall. DeepSeek and MiniMax still run through state-approved clouds. the weights may be open, yet the deployment layer stays locked.
GLM 5.2 looks like a very strong model, but I’m curious about the business model for open-source models. If companies can deploy the model themselves—or host it through providers like Fireworks or other clouds—how does the original model developer capture meaningful economic value? Just wondering what's the business model for those open source labs.
I agree the market has split into “as cheap as possible” vs “as clever as possible”
GLM 5.2 is clever, but to my taste DeepSeek writes better prose.
I wanted to like the Hermes models from Nous Research, but DeepSeek just beats them on all my private benchmarks.
I have belief in OM in directional future. But can’t really test it without company enable it to really do side to side comparisons.
It’s also harder to do so because the agentic workflow already tailored so much to CC setup, I don’t even get time to Codex even. It’s not like no time to work on codex, but I can’t wait to keep iterating on CC as it’s not done yet.
Personal projects most of sonnet level is pretty easy because of the complexity and context limits.
The open-source angle is neat, but framing it as a global leveller misses what happens inside China’s firewall. DeepSeek and MiniMax still run through state-approved clouds. the weights may be open, yet the deployment layer stays locked.
GLM 5.2 looks like a very strong model, but I’m curious about the business model for open-source models. If companies can deploy the model themselves—or host it through providers like Fireworks or other clouds—how does the original model developer capture meaningful economic value? Just wondering what's the business model for those open source labs.
Good question.
If Fireworks hosts GLM, Zhipu gets nothing. It only gets paid when:
1/ Customers go with Zhipu's own API
2/ On-prem (currently still Zhipu's main rev source)
My guess is they’ll keep open-sourcing strong models for distribution, but eventually keep the very best frontier model closed.
100% agree - see this coming.