Strands Decider 2B scores the options you give it, and then it stops

Strands Labs released Strands Decider 2B on October 1, 2026, a 2B decision model that chooses among options it is given and returns scores. TechCrunch's story is timestamped 9:49 AM PDT, and SiliconANGLE's update time is 12:00 EDT. The official post calls the release v19.

TL;DR - On Strands Decider 2B, the official post describes a Qwen3.5-2B torso, an LM head swapped for a pointer head of just over a million parameters, and rank-16 LoRA on the torso. There is no free-text output. - Official latency is a median of about 115 ms on a local RTX 3090, and about 153 ms for small tasks on an M3. On JevBench's public set it is 3rd of 33 in the 2B class. - SiliconANGLE says "v.20" and "less than 150 milliseconds." The official post says v19 and about 115 ms. TechCrunch carries a gold-rush caution from TypeSafe's CEO.

What actually happened

SiliconANGLE calls it v.20. Those labels disagree, and this piece uses the official one while keeping the conflict visible.

The shape of the model is the whole product pitch. Strands Decider 2B starts from a Qwen3.5-2B torso. The language-model head is replaced by a pointer head described as just over a million parameters.

The torso is fine-tuned with rank-16 LoRA. It does not answer in sentences. It picks among given options and returns scores. If you wanted a tiny writer, this is the wrong artifact, and the architecture is telling you that on purpose.

The official latency lines are specific about the box. Median latency is about 115 ms on a local RTX 3090. Small tasks land at about 153 ms on an M3.

"about" is doing work in both sentences. These are not presented as a cross-lab bakeoff. They are the release's own figures for those two machines.

On JevBench's public set, the release places the model 3rd of 33 in the 2B class. If just-over-2B models are excluded, it is 1st of 30. It scores 100% of the easy tasks.

That last clause is easy to over-read. One hundred percent of the easy slice is not a claim about the hard slice, and 3rd of 33 is not the same sentence as 1st of 30. The exclusion rule is doing the promotion. Say which denominator you mean.

Weights, training data, and scripts are on Hugging Face and GitHub, according to the release account used here. This piece does not add repo URLs that were not in the pack. "Open" in the tags sense means those artifacts were posted, not that a stranger has re-scored the leaderboard in these articles.

TechCrunch adds two human lines around the launch. Marc Brooker said it briefly hit the top Jevbench spot for its size before release. That is TechCrunch's report of what he said, not a block of quotation, and it sits next to a public-set rank that is 3rd of 33, not an ongoing first.

Before release and on the public set are different moments. Do not collapse them.

TechCrunch also quotes TypeSafe CEO Diogo Almeida. Read the line before you read the rank.

I get that people think it’s a gold rush, but they might be underestimating the difficulty of making the models actually smart.

The quote is a warning about the category, not a review of this checkpoint's easy-task score. A flood of decision models can still leave the hard part unsolved. Almeida is talking about that difficulty. He is not handing you a winner.

What to do as a reader

Use the official version label. The post says v19. SiliconANGLE says the release is "v.20" and describes latency as "less than 150 milliseconds." The official median is about 115 ms on a local RTX 3090, which is a different sentence from "less than 150," and the official M3 figure for small tasks is about 153 ms, which is not under 150.

They disagree. Do not average 115 and "less than 150" into a new number, and do not let a newsroom adjective overwrite the post the lab actually shipped.

When someone says it topped the board, ask which cut. Brooker, via TechCrunch, said it briefly hit the top spot for its size before release. The public-set line is 3rd of 33 in the 2B class, or 1st of 30 if just-over-2B models are set aside, with 100% of the easy tasks.

All three can be repeated accurately. They cannot be replaced by "it's the best 2B."

Judge it as a chooser, not as a chatbot. The job is: here are the options, here are the scores. If your workflow needs a paragraph, a tool call written out as prose, or an explanation, this release says no free text.

Wiring it into a slot that expects an LM head will waste a week. Wiring it into a slot that already has a closed menu is the experiment the architecture supports.

A fair local check, if you touch the weights, is to time your own menu on hardware you name, then compare that note with the official RTX 3090 median and the M3 small-task figure without pretending the machines match. Easy-task perfection is a reason to look at the easy set, not a reason to skip your own cases. Training data and scripts being public is an invitation to read them, not a finding that you already did.

Coverage framed the drop as an Amazon or AWS debut. TechCrunch's headline treats it as Amazon releasing a Jev-style model, and SiliconANGLE's headline says AWS debuts it. The release itself, in the account this piece follows, is Strands Labs shipping Strands Decider 2B. Keep the lab name on the version label you cite, which is v19, and treat the outlet headlines as their framing rather than a second version number.

What not to hype past the page

Decision models are showing up often enough that TechCrunch's piece is about a flood, and Almeida's line is the corrective. A pointer head, a small LoRA, and a fast median on one GPU do not answer whether the choices are wise. The gold-rush sentence is in the pack so you do not have to invent a mood. People can feel late and still be early to the hard part.

On Strands Decider 2B, nothing here is a reason to bet on a vendor, a cloud bill, or a token. The useful output of this story is a short spec: v19, no free text, official latencies tied to an RTX 3090 and an M3, a public-set rank with two denominators, and a newsroom that printed a different version string. If a slide has "v.20" and "under 150 ms" with no source, it is the SiliconANGLE wording, and it conflicts with the official post.

Read the official introduction, TechCrunch's account, and SiliconANGLE's account. Cite v19 unless you are explicitly quoting the outlet that disagrees.

Readers who want sourced recaps that are already on the site can read DogeOS public testnet and PUMP token buyback burn as separate live posts.