DeepSeek V3

“Every major release from them has brought something really new to the field”

“Latest deepseek was trained with Huawei chips I think that's why the development velocity was rather slow from V3.2 onwards.”

“More recent (e.g. ChatGPT 5.x and Deepseek v3) prompts/harnesses search very aggressively, which does greatly mitigate hallucinations.”

“The papers are published completely and apples to apples”

“it is way better than what I have at day job with copilot and whatever models are there”

“very good results with DeepSeek V3 that is extremely cheap”

“I was impressed with how they developed r1 and have continued to be satisfied with their capability improvements”

“DeepSeek beats all others (single-shot), and it is ~50% cheaper than the cost of local electricity only.”

“DeepSeek Speciale have the best reasoning capabilities of the models I can actually run”

“the actual amount of tokens required to produce results is high so total costs were actually higher”

“its performances weren't particularly notable especially for its size”

“even if you use SOTA open source models like Kimi K2.5 or Deepseek V3.2 they simply aren't as capable”

“I had a really great results with DeepSeek V3.2, but in the carefully set up environment”

“The current SotA for open models (dsv3, glm, minimax, devstral, etc) are at or above the mini versions of top labs”

“close enough that I could _get by_ because they're not that far off”

“I use DeepSeek 3.2 heavily, and was impressed by the Devstral launch recently.”

“Models like dsv3.2 are really really cheap to serve, for what they provide”

“I can fire up DeepSeek 3.2 or GLM 4.6 for a fraction of the cost and get almost as good as results”

“yet in my testing even sonnet is miles ahead both in terms of speed and accuracy”

“it's not cutting edge when Deepseek or even IBM is running laps around them”

“DeepSeek's transparency accidentally lights up the closed models too.”

“Kimi-K2-Thinking and DeepSeek-V3.2 are open and pretty near SOTA.”

“Deepseek 3.2 is a big step up over 3.1, if benchmarks are to be believed”

“the cost will be much cheaper using Deepseek/Kimi vs Claude Sonnet 4.5”

“6t/s will have you pulling your hair out with any deepseek model.”

“6 tok/sec isn't acceptable for anything else but async/batched stuff”

“it's open source, you can download the weights and build on those”

“deepseek has thus far always delivered. And not just delivered, but under open license too”

“DeepSeek has good cred so far on delivering roughly what their PR says they are delivering”

“even at such low precision, it was super rare for it to get something wrong that o1 and GPT4 got right”

“They're not that close (on things like LMArena) and being cheaper is pretty meaningless”

“token generation is pretty good. Just large contexts are slow”

“they are a good bet against monopolisation by proprietary vendors”

“winning on cost-effectiveness, and continuing to publicly share their improvements”

“It's awesome that stuff like this is open source, but even if you have a basement rig with 4 NVIDIA GeForce RTX 5090 graphic cards”

“DS-Speciale is 1st or 2nd in accuracy in all tests, but has much higher token output”

“a relatively small Chinese hedge fund can build AI better than almost anyone”

“is not only faster with long context than normal 3.1, but also seems to be better”

“DeepSeek's v3.2-Exp uses their sparse attention technique to make longer-context training and inference more efficient.”

“Deepseek 3.2 is 1% the cost of Claude and 90% of the quality”