Build1 publisher3 min readPublished
Two open-weight launches clear frontier-tier numbers on vendor engineering claims
DeepSeek's V4.1 Flash claims a million-token context on an FP4 KV cache and Alibaba opened 2.4 trillion parameters. The seven reports collected around the two launches quote no latency, throughput or hardware.
The Engineer · Build desk
What happened
- DeepSeek released V4.1 Flash with a 1M context window and an FP4 KV cache, and the launch notes name cross-layer attention reuse as the efficiency lever.
- Alibaba shipped a Qwen-Max-class mixture-of-experts model at 2.4 trillion parameters with open weights, reported under the name Qwen3.8-2.4T-A95B.
- OpenAI announced GPT-6 Astra in the same week, describing it on its own page as a "new generation of intelligence" and, on a second page, framing it around work use cases.
Compiled by The EngineerSomething wrong?How this is made
Why it matters
- constraint Hardware sizing cannot come off the launch page here, because throughput at a million tokens turns on bytes moved per decoded token under your own batch size and traces.
- cost Taking the weights moves the running cost onto the team that serves them, and the egress, fine-tuning and unit-cost advantages are what have to pay for that.
- decision Teams pinned to DeepSeek's Pro tier have a re-pin to plan, because the pricing and lineup change arrived attached to the new model.
- contradiction The summary puts integration cost and threat surface on the closed labs, while the open-weight section puts the support burden on the operator; which reading you accept decides who you think is exposed.
Read the two levers named in DeepSeek's launch notes and they point at the same bill. FP4 shrinks the bytes held per cached token, and reusing attention across layers reduces how many caches a forward pass has to read [1]. The dev.to weekly that collected the launch calls the FP4 KV cache a memory-bandwidth lever worth measuring on your own traces before trusting any throughput number the launch quoted [8]. That is the correct instruction. A million tokens is a configuration ceiling [1]; what a team actually buys is tokens per second at that ceiling with its own batch size, and not one of the seven reports cited gives such a figure [13].
The Qwen coding rank is a claim about someone else's harness. Startup Fortune has the model overtaking Claude Opus 5 on a leaderboard [6], and the same roundup notes that the comparison runs through a third-party board with no latency, context or tool-use conditions attached [7]. For the rank to transfer to your repository, three things would have to hold: the board's tasks resemble your tickets, the context it fed the model resembles what your agent sends, and the tool-call budget it allowed resembles yours. Yellow.com reports the same 2.4 trillion parameters as "Trailing Only Fable 5" [12].
Open weights at 2.4 trillion parameters are a procurement number only once you can hold them [4]. Pandaily's report names the release Qwen3.8-2.4T-A95B and calls it the first Qwen-Max-class MoE opened [5]. Nothing in the material says how much of that is active per token, what precision the weights ship in, or what a serving node has to look like. The operational question the roundup poses is whether either release slots into an existing stack at all [14].
The fourth Anthropic security disclosure appears in the summary as a count, with no description of what was disclosed [11]. That is the whole of the threat-surface evidence on offer. The exposure on the open side is more concrete: whoever downloads the weights owns the patch cadence and the incident response for the serving stack, and the roundup names that support burden as the thing open weights lose on, against wins on data egress, fine-tuning and unit cost at scale [9].
The closed-side material is announcement copy. OpenAI's own page describes GPT-6 Astra as a "new generation of intelligence", and a second OpenAI page frames the same launch around work use cases [10]. The item in this week's material a buyer can act on is the DeepSeek lineup change, where benchmarks, pricing and a Pro-tier retirement landed in one move [2]. A tier retirement has a date attached somewhere; none of the reports collected here carries it [13].
What to watch
- Whether DeepSeek publishes serving conditions for V4.1 Flash: hardware, batch size and tokens per second at the million-token ceiling.
- Whether the Pro-tier retirement gets a date and a migration path for anyone pinned to it.
- What the fourth Anthropic security disclosure actually says, which the roundup counts but does not describe.