Naderu Weekly · Episode 01
The open-weights inversion
At midnight UTC on Monday, Moonshot published the weights of a 2.8-trillion-parameter model — reported as the largest open-weight release so far. Anyone can download it. Almost nobody can run it. In the same month, Meta released its first paid closed model. This is what that pair of facts actually changes.
Ira and Nitin are synthetic voices; the research and the arithmetic are ours, and every figure below is linked to where we got it. If you'd rather have it arrive than remember to check, subscribe to the weekly report — one email a week, one click to leave.
Six things that moved
| What | Detail | Sourcing |
|---|---|---|
| Kimi K3 weights published | 2.8T parameters, sparse mixture-of-experts, ~1.4 TB on disk at MXFP4, 1M-token context, native text, image and video. 00:00 UTC, 27 July. | second-hand |
| Meta ships a paid closed model | Reported as Meta's first paid closed release, with focus shifting to monetising the inference layer rather than publishing weights. | second-hand |
| Output prices collapse | Three flagship-tier releases inside roughly 24 hours; new-tier output reported around $4–$6 per 1M against $25–$50 for the previous generation. | second-hand |
| Claude Opus 5 | Released 24 July at $5 in / $25 out per 1M — half of Fable 5 — and top of the public leaderboards on release. | first-party pricing |
| GPT-5.5 reasoning tiers deprecated | Marked deprecated while still listed at $5 / $30. | second-hand |
| DeepSeek V4 reaches GA | General availability, retiring the previous API in the process. | second-hand |
Where a row says second-hand, we could not find the claim on the vendor's own page this week. It is still reported here, but labelled — and said out loud in the episode.
Two directions at once
Put the first two headlines side by side and the shape of the year appears. The open frontier has moved east: Moonshot with Kimi K3, DeepSeek's V4 line under MIT, Z.ai's GLM-5 line under MIT, Alibaba's open Qwen models under Apache-2.0. Meanwhile the company that did most to make open weights a Western default has moved the other way.
Both can be rational at the same time, which is the part worth sitting with. Publishing weights is how you buy relevance when you are not the default. Withholding them is how you protect margin when you are. Neither is a moral position; both are strategy.
Open is not the same as runnable
Here is where the celebration usually outruns the engineering. Take the figure as reported: about 1.4 TB of weights. That is not a download problem, it is a residency problem — the weights have to sit in accelerator memory before the model emits a single token. On 80 GB cards that is eighteen of them for the weights alone. That is arithmetic on the published number, not a vendor claim.
Then you still need room for the key-value cache, and this model advertises a one-million-token context. Long contexts are exactly where that cache grows. You are into a multi-node deployment with a fast interconnect before you have served one user.
This is the most common misreading we encounter, so it is worth stating plainly. In a sparse mixture-of-experts only a few experts fire per token, so you pay for fewer floating-point operations. But you cannot predict which experts the next token will want, so all of them must stay resident. The efficiency is real and it is entirely on the compute axis.
And notice what the same week handed you as an alternative: that model's hosted API runs around $3 in and $15 out per 1M with cached input near $0.30, and further down-market DeepSeek's own pricing page lists a flash tier at $0.14 in and $0.28 out. For most teams, renting is not a compromise. It is simply correct.
The question underneath: what do you need custody of?
The argument was never really open versus closed. It is a question about custody, and the answer usually falls in one of three places.
Regulated data, air-gapped networks, a review that will not accept egress to anyone's API. That is a real constraint and it ends the conversation. But notice it pushes you toward a model you can genuinely serve, which is almost never the largest one on the leaderboard.
It has to answer the same way in two years, on your vocabulary, with your edge cases. That is a fine-tuning and evaluation problem, and it wants a model small enough to train and pin — with the recipe and the eval suite kept under version control so the result is reproducible.
You need a capability this quarter and the model is not your product. Rent it. The price war is genuinely on your side, and anyone telling you to buy accelerators for that is selling accelerators.
The catch in renting: a rented price is a dated price
This week put the trap on display. A model marked deprecated while still on the price list is a migration scheduled by somebody else. And it is not only retirements — some of the most consequential numbers are published in advance and simply not read:
| What | When | Effect |
|---|---|---|
| Claude Sonnet 5 introductory pricing ends | 31 Aug 2026 | $2 / $10 becomes $3 / $15 — a 50% rise, on the calendar, in a public document |
| Gemini and Grok context step | Above 200K in | Per-token price doubles, so the large window costs twice as much to fill |
| Claude 4.7+ tokenizer | Already | ~30% more tokens for the same text, making like-for-like rate comparisons wrong by about a third |
That last one deserves emphasis because it is invisible on every pricing page. A rate card compares dollars per token; it cannot tell you that the tokens changed size. It is why we price a workload rather than a model — your real prompt, your real cache hit rate, your real output length, compared as invoices. We have cases on the model board where the cheaper model per token is the dearer one per job, purely because one provider caches properly and the other does not.
Three things to take away
- Price the workload, not the model. The rate card is the smallest term in the sum.
- Decide what you need custody of — weights, behaviour, or neither — before you have the open-versus-closed argument. It settles most of it for you.
- Pin model versions and diary the deprecations. This week alone gave us a deprecated tier still on sale and a price rise dated 31 August. None of it arrives by email.
Sources
Every figure above, in the order it appears. first-party means the vendor's own page; second-hand means we could not find it there this week and are relying on a report.
| Source | Kind |
|---|---|
| Naderu model board | first-party |
| Anthropic pricing docs | first-party |
| Gemini API pricing | first-party |
| xAI models & pricing | first-party |
| DeepSeek API pricing | first-party |
| Together AI pricing | first-party |
| Kimi K3 vs DeepSeek V4 Pro vs GLM-5.2 — licences and serving cost | second-hand |
| Kimi K3 is live — pricing, benchmarks, and the wait for public weights | second-hand |
| AI news roundup, 26 July 2026 | second-hand |
| Platform Watch — the model price war | second-hand |
| Best AI models, July 2026 | second-hand |
| DeepSeek V4 hits GA, old API retired | second-hand |
Ira and Nitin are synthetic voices. The research, the script and the arithmetic are the Naderu team's, and every figure is linked above so you can check us.
Naderu is an AI-models company and a venture of BytesBrains Pte. Ltd. We train, release and run specialised models. We are model-agnostic by policy and take no payment for placement on the board — nobody can buy a row or a rating.