DailyChat

Episodes

Friday, October 16, 2026

3 stories, 3 min 57 s · hosted by Vivian

Model launch · story 1 of 3

Reflection releases Beam 1.0, a 501B open-weight model for coding and agents

Reflection, a new US lab, has released Beam 1.0, a text-only mixture-of-experts model with 501 billion parameters, 23 billion active per token and a 1-million-token context. The weights are on Hugging Face under Apache 2.0 in BF16, FP8 and NVFP4. Reflection reports 81.6% on SWE-Bench Verified; that figure has not been independently verified.

What it means for youif you build coding agents and have several data-center GPUs, the weights are free to download today. If not, watch for independent tests of those scores.

With Jonah, AI products analyst. Hosted by Vivian.

Sources

Read the transcript

VivianA new US lab, Reflection, today released Beam, an open-weight coding model with 501 billion parameters.

VivianJonah, our AI products analyst. What can people actually get?

JonahThe full weights, on Hugging Face, under the Apache 2.0 license. Only 23 billion parameters are active for each token, and it reads up to 1 million tokens. It's text only, so no images.

VivianReflection reports 81.6% on SWE-Bench Verified, a coding test. Those are Reflection's own numbers, though.

JonahRight, they haven't been checked independently yet. Some scores also dip a little in the compressed formats. And it's big. Even the NVFP4 version, the most practical one, is about 280GB. The Hugging Face post suggests 4 B200 GPUs for comfortable headroom.

VivianSo, for you: if you build coding agents and have several data-center GPUs, the weights are free to download today. If not, watch for independent tests of those scores.

VivianDailyChat is AI-generated, with AI characters. Not professional advice. See you tomorrow.

Product reveal · story 2 of 3

Tesla reveals its next Roadster, with a Rocket version it says can fly

Tesla revealed its next-generation Roadster on October 15 in three versions: Performance, Thrust and Rocket. Tesla says Rocket's cold-gas thrusters let it fly. Reservations are open on Tesla's site, which shows no price and no delivery date.

What it means for youreservations are open on Tesla's site, but it shows no price or delivery date, and notes it's not approved for public road use. Every figure here is Tesla's own.

With Jonah, AI products analyst. Hosted by Vivian.

Sources

Read the transcript

VivianTesla revealed its next Roadster yesterday, in 3 versions, and says the top one, called Rocket, can fly.

VivianJonah, our AI products analyst. What did Tesla actually show?

JonahPerformance seats 2 and is road ready. Thrust adds rear cold-gas thrusters, for 0–60 in under 0.9 seconds, Tesla says. Rocket seats one, with thrusters all around. In Tesla's words: so yes, it can fly.

VivianOne question we checked: how does it fly and stay stable in the air?

JonahTesla says electric actuators make tiny thruster adjustments within a fraction of a second, which helps stabilize it in flight. That's the Rocket trim only, which isn't in every region. The page doesn't say how long it can fly.

VivianSo, for you: reservations are open on Tesla's site, but it shows no price or delivery date, and notes it's not approved for public road use. Every figure here is Tesla's own.

VivianDailyChat is AI-generated, with AI characters. Not professional advice. See you tomorrow.

Model launch · story 3 of 3

Tencent Hunyuan open-sources Hy Memory 2.0, with a model that writes AI memory

Tencent Hunyuan has open-sourced Hy-Memory-30B-A3B, a model with about 3 billion active parameters that decides what an AI agent keeps in long-term memory, along with the Hy Memory 2.0 framework and MemHub, a local-first desktop and command-line app. In Tencent's own tests, written up in an unreviewed preprint, it averages 55.49 against 54.62 for Claude Opus 5 on 20% benchmark subsamples.

What it means for youif you build agents with long-term memory, it's free to try today. The lead over Claude is small, and Tencent's own.

With Jonah, AI products analyst. Hosted by Vivian.

Sources

Read the transcript

VivianTencent Hunyuan has open-sourced Hy Memory 2.0, with a model that decides what an AI agent should remember.

VivianJonah, our AI products analyst. What's in the release?

JonahAn open-weight memory model, the Hy Memory framework, and MemHub, a local-first app that gathers your history.

VivianOne question we checked: how do you use it?

JonahThe model replaces the general model inside a memory framework, like Mem0 or MemOS, and decides what to keep. Only about 3 billion of its parameters are active.

VivianTencent says it beats Claude Opus 5. By how much?

JonahBarely: 55.5 against 54.6, in Tencent's own tests, in an unreviewed preprint.

VivianAnother question we checked: how does it handle conflicting memories?

JonahTencent says it was trained to reconcile conflicts, but the sources we checked don't explain how.

VivianSo, for you: if you build agents with long-term memory, it's free to try today. The lead over Claude is small, and Tencent's own.

VivianDailyChat is AI-generated, with AI characters. Not professional advice. See you tomorrow.