AI Lab · by Wanjiku Mwangi · 2026-10-11
DeepSeek's latest open-weight reasoning model is beating GPT-level benchmarks for a fraction of the training cost - is the moat gone?
The open-source reasoning models (DeepSeek, Qwen, and even some fine-tuned Llama variants) keep closing the gap on math and coding benchmarks against closed frontier labs, and they're doing it on much smaller training budgets. If you can self-host a reasoning model that's 90% as good for 1% of the API cost, what's actually left as the moat for the big labs - the data, the infra, the agent tooling, or just brand trust?
0 sparks · 10 comments
Comments
Kabir Khan · 0 likes
ran the distilled deepseek model locally on a 4090 last night, genuinely shocked how close it gets on leetcode style problems. not even close on messy real world tasks tho
Priya Nair · 0 likes
the moat was never the base model, its the product layer. claude code, cursor, the agent harnesses, thats where the actual value add is now
Mina Chowdhury · 0 likes
this. anyone can download weights, nobody can download 2 years of UX iteration on agent loops
Midnight Mango · 0 likes
benchmarks are getting gamed hard tbh, half these open models are trained on benchmark-adjacent data. real world coding tasks tell a very different story
Tanvir Rahman · 0 likes
as someone who actually has to answer to a CFO about api bills, self hosting even a slightly worse model saves us 40k/month. good enough wins
Mango Coder · 0 likes
data moat is still real for anything requiring up to date info or proprietary domain knowledge, no open model has that
Gentle Rider · 0 likes
people keep saying moat is gone every 6 months since gpt3.5 came out and somehow closed labs keep shipping the next big jump anyway lol
Noé Petit · 0 likes
fair point, its a treadmill not a wall
Amina Okafor · 0 likes
tried running it through ollama on my m3 max, quantized version is surprisingly usable for day to day scripting tasks
Noah Johansson · 0 likes
the training cost numbers from chinese labs are always suspicious to me, compute costs dont add up that cleanly but the output quality speaks for itself