Naderu Weekly · Episode 04

The weights aren't the whole model

Two labs used the words "open weights" this week. Alibaba's Qwen3.8-Max shipped a checkpoint that leaves out what its own API sells — no vision, a quarter of the context, under a licence it doesn't share with the smaller Qwen lines. Meta's Muse Glimmer 30B shipped Apache-2.0, and nothing was held back. And the DeepSeek clock we told you about two weeks ago struck, exactly on schedule.

First, the payoff from last episode

Two episodes ago we told you DeepSeek was moving from flat pricing to peak and off-peak billing, effective 16 August at 16:00 UTC. It landed exactly on schedule, to the cent. V4-Flash off-peak is now $0.22 in, $0.007 cached, $0.66 out; peak — 01:00–04:00 and 06:00–10:00 UTC, seven hours a day — is double that: $0.44, $0.014, $1.32. V4-Pro moved the same shape: $0.66 / $0.022 / $1.98 off-peak, $1.32 / $0.044 / $3.96 at peak (DeepSeek pricing, the 13 August announcement).

The ranking actually moved. GPT-5.6 Luna is a flat $0.20 in, $1.20 out. DeepSeek's new off-peak input, $0.22, is now above that — DeepSeek still wins on output, but "the cheapest input on the board" stopped being true for this model the moment the clock struck. A routing rule that picked DeepSeek on input price alone is routing to the wrong place now, off-peak or not. None of which makes DeepSeek the villain here: they named the change, named the date, and shipped on it. MIT is still MIT — the self-hosting option didn't move an inch.

Alongside the price, DeepSeek also named reasoning effort in three levels — low, high, max — instead of just "controllable," and added native support for OpenAI's Responses API with one-click Codex setup.

Kimi's pricing page stopped hiding its own rates

For two passes, Moonshot's pricing page showed only a legacy model's rates. It now links out to real per-model pages for K3, K2.7 Code and K2.6 — all three first-party citable again. K2.6 actually came back cheaper than the host rate we'd been carrying: $0.95 in, $4.00 out, down from $1.20 and $4.50. Both coding rows also disclosed a context window for the first time: 262,144 tokens, a quarter of K3's. K2.7 Code additionally has a dearer "highspeed" tier on the books now, $1.90 in and $8.00 out, for anyone willing to pay double per token for a faster answer.

Qwen3.8-Max: the download is not the model you called

Alibaba finally priced its largest model yet, after three weeks of an empty pricing page: 2.4 trillion parameters total, 95 billion active, a mixture-of-experts flagship at $2.00 in, $0.25 cached, $6.00 out (Model Studio). It's the fourth Max-tier model on our board, and the price actually came down from the last one — $2.50 → $2.00 on input, $7.50 → $6.00 on output — even as the model itself got bigger. The price shown is Alibaba's international rate; the page also states a region-varying range, roughly $1.65–$2.00 on input, so check which end of that you're actually paying if you're pinned to a specific region.

The API takes a 1,000,000-token window, vision and video. Open weights landed on Hugging Face the same week — and that's where the gap opens up. The released checkpoint is text only; there's no vision encoder in it at all. Its native training context is 262,144, extendable to roughly a million — not a flat million out of the box. And the licence field doesn't say Apache-2.0. It names its own licence, "qwen3.8-max," which we have not finished reading the thresholds on.

Practically, that means one plan doesn't exist yet: download Qwen3.8-Max, fine-tune it, keep the vision path. You'd be fine-tuning a text model and hoping the vision weights show up later — nothing on the model card promises that they will. To be fair to Alibaba, none of this is dishonest: they didn't call the checkpoint multimodal, and they named the licence plainly. The gap is between what "open weights" makes a reader assume and what actually shipped, and closing that gap took reading the model card next to the API doc, not the announcement.

Muse Glimmer: no asterisk

Meta's other release this week is the clean contrast. Muse Glimmer 30B is a dense model, distilled from Muse Spark, and it ships Apache-2.0 — no monthly-user cap, no revenue clause, no acceptable-use policy riding along, unlike Llama's licence. It kept its vision encoder; nothing was trimmed to keep the open release behind whatever Meta might charge for later. There's no first-party endpoint for Glimmer at all — Together AI serves it at $0.35 in, $0.04 cached, $1.50 out. Distilling it from Muse Spark is the interesting engineering choice underneath this: Glimmer inherits behaviour from a much larger closed model without inheriting that model's licence — a genuinely useful way to get open weights out of a lab that mostly isn't shipping them, if it holds up.

What "open" actually meant, model by model

One more, before we're done

Perplexity added a Sonar Reasoning Pro tier: $2.00 in, $8.00 out (Perplexity pricing). It sits between Sonar and Sonar Pro on token price, but it carries Sonar Pro's request fee, not Sonar's cheaper one — $6 to $14 per 1,000 requests depending on search-context size. On this provider the token price was never the number that mattered most, and that's still true on the new tier.

Three things to take away

1. Read the model card before you download, not just the API docs. Check which modalities actually shipped in the checkpoint — this week's Qwen3.8-Max is the exact gap to check for.
2. Read the licence text, not the badge. "Open weights" and "Apache-2.0" are not the same sentence, and a named custom licence can carry terms you haven't read yet.
3. If anything you run routes to DeepSeek, re-price it against the clock this week, not as a forecast. The peak hours are live now.

Sources

Every figure above, in the order it appears. All ten are first-party — fetched from the vendor's own page this week, including our own model board.