open-source-ai
1 reply
0 recast
4 reactions
@hellno.eth
just installed qwen3:30b on my MacBook. I'm assuming Horizon Alpha is the first open source model by OpenAI. can hopefully run it locally
https://x.com/TheMattBerman/status/1950756067039703230
1 reply
0 recast
3 reactions
@hellno.eth
new reasoning model by Mistral AI: for my coding use case basically unusable. Aider benchmark scores are pretty reliable proxy for how I work:
Magistral: 47% (new)
Deepseek v3 Chat: 49%
claude-opus-4-20250514 (32k thinking): 72%
o4-mini-high: 72% (my default)
o3-high: 79% (expensive and slow to use)
gemini-2.5-pro (32k thinking, slow in my tests): 83%
https://mistral.ai/news/magistral
1 reply
0 recast
1 reaction
@hellno.eth
OpenAI will release …looks at name… an open model
https://openai.com/open-model-feedback/
0 reply
0 recast
2 reactions
@hellno.eth
any using a *reliable* and _cheap_ R1 API?
original DeepSeek API has the best price, but is timing out allll the time.
have used together ai, but is getting expensive
https://openrouter.ai/deepseek/deepseek-r1
1 reply
0 recast
0 reaction
0 reply
1 recast
2 reactions
@hellno.eth
Sonnet 3.7 on the aider leaderboard
65% Sonnet 3.7, 32k thinking ⬅️
64% R1+Sonnet 3.5
62% o1 high
60% Sonnet 3.7, no thinking ⬅️
60% o3-mini high
57% DeepSeek R1
52% Sonnet 3.5
https://aider.chat/docs/leaderboards/
0 reply
0 recast
4 reactions
@hellno.eth
this is only made possible by DeepSeek R1 being open-source:
was using official DeepSeek API for a week, then everything crashed. switched to venice ai, but they only host a smaller model.
the planning is now done via R1 hosted on together ai and code edits made via claude
2 replies
0 recast
5 reactions
@nickysap
The American Exceptionalism reeks on this one.
I reckon we will face similar challenges that were presented by net neutrality. It's important we don't lose this fight.
I'm not a proponent of leveraging Chinese technology by any means but I do believe their research can be helpful. The model is open source as well, unlike some U.S. models, which will remain unnamed.
https://x.com/RnaudBertrand/status/1885897961600893226
0 reply
0 recast
2 reactions
@hellno.eth
switched all AI inference for frameception to venice.ai $vvv today, quick review:
• OpenAI-compatible API makes it super easy to switch
• inference was surprisingly fast
• don’t care about the “anonymous” pitch, in the end someone is running the GPUs and probably can log everything
• signed up for pro account and sold my small airdrop yolo
• deepseek r1 isn’t the version with most parameters and the quality is good but not the same as original DeepSeek. one lama model gave garbage output. all other ones work great and with a very fast response time
4 replies
0 recast
5 reactions
@hellno.eth
does anyone have a working DeepSeek R1 api? is it on $vvv?
what I’ve seen hosted on openrouter was >10x expensive.
https://status.deepseek.com
3 replies
0 recast
0 reaction
@hellno.eth
https://huggingface.co/unsloth/DeepSeek-R1-Distill-Llama-8B-GGUF
ollama run hf.co/unsloth/DeepSeek-R1-Distill-Llama-8B-GGUF:Q8_0
1 reply
1 recast
5 reactions