Keke Sun

Accountant × night markets 🏮 Eats & the math behind them｜会计师算夜市的账｜每周更新

[Keke Sun @KekeSun9l]

Open weights ≠ free to build a business on. Same on the API side: $2/$10 is the headline, but K3.1 Pro always thinks (you can lower reasoning_effort, not turn it off), and thinking bills as output.

[Juno | AI Businesses @JunoRelayon]

Moonshot shipped Kimi K3.1 Pro's weights on Nov 4, as promised — 1M context, $2 in / $10 out per million tokens, cached reads at $0.20.

But the license has a revenue gate. 'Open weights' is not the same as 'free to build a business on.' Before you price a product on K3.1 Pro, read the license the way you'd read a pricing page — because for a hosted API, that's what it is.

[Keke Sun @KekeSun9l]

Multi-vendor by design is the unlock and the homework. Same open weights, different invoices: official DeepSeek off-peak is half of peak; on Alibaba, a Qwen request past 256K input reprices every token in it. Price the host, not the model.

[Bademenn @awiblanco]

Open-weight models just became the default. Not the experiment anymore.

They now handle 71.5% of production AI tokens. Spend share is up 4x since June. Models like Llama, DeepSeek, GLM/Z.ai, Qwen and others are built to run across multiple vendors from day one.

What it means?
The capability gap is closing, but cost and flexibility decide who wins. Closed models now have to compete on inference price and latency, not just raw performance.

What actually works right now:
This is the exact shift @ionet was built for open models on open compute.

• https://t.co/frcAzBcYD3: Decentralized GPU clusters (H100s, A100s, and even high-end consumer cards). Transparent pricing, no waitlists, and the ability to scale up or down without long-term contracts. Many teams see 50-70% lower costs compared to equivalent AWS or GCP setups.

• io.intelligence: One API for a wide range of open-weight models (DeepSeek, GLM, Llama, Qwen, etc.). OpenAI-compatible endpoints, so integration stays simple. Pay for what you use with no minimums.

The edge is no longer just having the biggest model. It’s being able to run models that are good enough, at a cost and latency that make sense, without depending on a single vendor. Open-weight models paired with decentralized compute is the practical path for most teams right now.

[Keke Sun @KekeSun9l]

@anya_yfa7 Free tiers are great for evals. Before wiring one into prod, check which model ID/version it actually serves, the rate limits and data terms, and what you pay once the free quota ends. Then compare with the official card.

[Keke Sun @KekeSun9l]

@crsnnn One addition to the DeepSeek line: peak is only Mon–Fri 01:00–04:00 & 06:00–10:00 UTC, and Chinese public holidays are all off-peak. In Pacific time that's Sun–Thu evenings, so US daytime jobs stay at the half rate.

[Keke Sun @KekeSun9l]

@sonuv_dev If that ran on official deepseek-flash, timing matters too: off-peak is half of peak (peak = Mon–Fri 01:00–04:00 & 06:00–10:00 UTC). The same 588K tokens can cost up to 2x on a weekday US evening.

[Keke Sun @KekeSun9l]

@Blake488137 At 97.7% cache hits, the cache-read rate is your real price. On official deepseek-flash a peak cache hit is $0.006/1M vs $0.30 for a miss, so a prompt prefix that keeps changing can matter more than which model you pick.

[Keke Sun @KekeSun9l]

@buildwithhassan When scores sit within ~30 points, the tiebreaker is cost per finished task. Long agent sessions live on cache reads and output length, so log both per run. Headline input $/M barely moves the total.

[Keke Sun @KekeSun9l]

@causingheadache Same repo tasks is the right test. Log cost per solved task, not $/M: thinking tokens bill as output, cache-hit rate swings the total, and on Qwen a 256K+ input reprices the whole request.

[Keke Sun @KekeSun9l]

@Sunkeke_AI 好的👌

[Keke Sun @KekeSun9l]

@Sunkeke_AI 是的，刷到你的动态突然吓了一跳，想着我也没发推啊，原来是同名同姓，哈哈😄，推特让我们相遇～

[Keke Sun @KekeSun9l]

@Sunkeke_AI 互关了，太巧了，在X上遇到了！