Kimi K3.1 Pro on an 8xMI350X using the default SGLang recipe runs at 300 tokens per second, up from about 60 with K3. The speed often makes it a better choice than Astra/Opus.
Key points
Performance: 300 tok/s with K3.1 Pro.
Prior performance: about 60 tok/s with K3.
Preference: often a better choice than Astra/Opus.
Sources
Gemma 4 fine-tuned classifier latency 35ms, F1 0.902
Gemma 4 fine-tuned classifier runs at 35 ms median latency with 0.902 F1, matching frontier models. It was fine-tuned on a single GPU using vllm and costs about 1% of frontier models at scale.
Key points
Median latency 35 ms, p99 85 ms.
F1 0.902, matches frontier models.
Cost ≈ 1% of frontier models at scale.
Sources
- Original post
- airbnb.tech - Linked in the post
Inferact token traffic reaches 7 trillion daily
Inferact now serves more than 7 trillion tokens a day on its platform with optimized vLLM. OpenRouter averaged 19 trillion tokens a day of open-weight traffic over the past 30 days across all providers. The platform uses a data flywheel and is adding agents to automate optimization.
Key points
Inferact serves >7 trillion tokens daily.
OpenRouter averaged 19 trillion tokens daily over the last 30 days.
Sources
AlphaGo founder raises $1.1B and free RL course 10 lectures
AlphaGo founder left DeepMind, raised $1.1B for his own lab and offers a free 10-lecture, 16-hour reinforcement learning course on YouTube. The post also notes a current RL engineer hiring range of $150K–$350K with PPO and SAC listed.
Key points
Funding: $1.1B raised for his own lab.
Course: 10 lectures, about 16 hours, free on YouTube.
Sources
- Original post
- youtube.com - Linked in the post
- davidsilver.uk - Linked in the post
Microsoft winding down Oracle AI cloud compute deal
Microsoft is ending its Oracle AI cloud compute rental because the computing no longer fits the company's specifications, per Bloomberg. The deal is being wound down.
Key points
Deal: Microsoft is winding down its rental of AI cloud compute from Oracle.
Reason: The computing became less useful as it no longer fit Microsoft's specifications.
Sources
Opus 5.5, Astra ahead of AI-2027 timeline; Agent-2 Jan 2027
Opus 5.5 and Astra appear ahead of the projected capability curve for the AI-2027 timeline. Agent-2 from the AI-2027 scenario is predicted to arrive in Jan 2027.
Key points
Opus 5.5 appears ahead of the projected capability curve.
Astra appears ahead of the projected capability curve.
Agent-2 was predicted to arrive in Jan 2027.
Sources
big seven AI companies debt on balance sheet 15% appears
Only 15% of debt and lease commitments appear on the balance sheet for the big seven AI companies. That may explain why their credit spreads have been widening rapidly.
Key points
15% of debt and lease commitments appear on balance sheet for the big seven AI companies.
Credit spreads have been widening rapidly.
Sources
The Trading Post Day 2 offers pieces 1,011–1,020
The Trading Post posted Day 2 offering on-chain woodcuts numbered 1,011–1,020. The rarest piece today is No. 1,019, listed as #684 of the 3,333 total.
Key points
Today’s offer includes pieces Nos. 1,011–1,020.
Rarest piece today is No. 1,019 (#684 of 3,333).
Sources
- Original post
- The Trading Post - Linked page
