Meituan releases weights for its 1.6-trillion-parameter agent model
Meituan's LongCat-2.5 weights are now public, trained entirely on Chinese hardware and close to Claude Opus 5 on terminal tasks. Elsewhere, Starship was caught by its tower and Meta put out a cheaper Muse Ember.
Koda editorial team · 16 November 2026 · 5 min read
Meituan showed a 1.6-trillion-parameter agent model running on 194 Huawei chips, and Microsoft's hourslong global outage, which it traced to an earlier network capacity upgrade, is a reminder that where models run matters as much as how well they score.
Today: Build your stack so any model can be swapped out. The steps
Meituan's LongCat team released the weights for LongCat-2.5 on November 9, 2026, a week after unveiling the 1.6-trillion-parameter agent model. It scores 88.4 on Terminal-Bench v2.1, slightly behind Anthropic's Claude Opus 5. It was trained entirely on Chinese hardware, and a sample deployment runs on 194 Huawei chips.
Why it matters. Developers can now download an agent model that nearly matches Anthropic's best on terminal tasks and run it on Huawei chips, instead of paying for access to a closed rival.
Bland launched Tofu-1, a research-preview full-duplex speech-to-speech model that listens and speaks at once on phone calls.
Why it matters. Full duplex lets a phone agent handle interruptions mid-sentence, which removes the turn-taking lag that drives callers to hang up on scripted voice bots.
SpaceX's Starship completed a 22-hour orbital trip, flew back to Texas and was caught by the mechanical arms of the tower it launched from. It is the first time SpaceX has done this with Starship.
Agility, the maker of Digit, listed on Nasdaq at a $2.5B1 pre-money valuation despite only $1.8M in revenue last year, becoming America's first publicly traded pure-play humanoid company. The listing gives investors a direct bet on humanoids and exposes the company to quarterly scrutiny.
Meta's Cheaper Muse Ember
Reports describe Meta's Muse Ember as substantially cheaper than comparable models and competitive with Claude Haiku on an intelligence benchmark. It joins Muse Wonder, released October 12, and Muse Spark 1.1, which Meta offers through its model API, putting Meta in competition on the price of small models.
NATO will skip next year's leaders' summit in Albania and cut back high-level meetings. Allies are split over defense spending, Iran and the war in Ukraine.
Why it matters. Defense suppliers counting on summit-driven spending pledges should expect slower, less coordinated contract signals from NATO allies through next year.
S&P 500 -0.40%, the Nasdaq -0.91% and WTI crude -0.20% on the 13 Nov session. Bitcoin +1.32% and Ether +2.86% on the 16 Nov session. The crypto Fear & Greed index reads 10, Extreme Fear.
S&P 500
7,561.67
-0.40%
Nasdaq
25,136.27
-0.91%
Bitcoin
$64,990.37
+1.32%
Ethereum
$1,819.80
+2.86%
Crude Oil (WTI)
$88.56
-0.20%
Crypto Fear & Greed
10 / 100
Extreme Fear
Levels from Yahoo Finance2, 13 Nov and 16 Nov; alternative.me3, 16 Nov.
Go deeper
The Koda Signal, 16 November 2026
Meituan releases weights for its 1.6-trillion-parameter agent model
Insight7 runs two products on one site: self-serve call scoring with metered allowances you can buy today, and AI voice and chat agents you reach through a demo booking. The scoring side is a reasonable bet for support and sales managers who want each conversation graded against their own rubric and coached from the gaps. The agent side has too little public evidence to buy on faith, so keep it behind a simulation gate until it earns live traffic.
Pick models by testing them and keep the switch cheap.
With LongCat-2.5 weights downloadable and reports of Meta's Muse Ember undercutting comparable models on cost, the model you choose this month may not be the one you run next quarter. The 15.4-point slide on Terminal-Bench v4.0 shows why your own hard tasks should make the call, not a single leaderboard number. Keep an adapter in front of every model and a small private eval set ready, then rerun it whenever a new release lands.
1Put every model behind one adapter. Route calls through a single interface so LongCat-2.5 can sit next to a closed model in the same app. The MIT license means you can self-host the weights and compare them side by side.
2Test on your own hardest tasks. A 15.4-point drop between Terminal-Bench v2.1 and v4.0 shows an easy suite can hide real gaps. Run your toughest terminal workflows against any candidate before you switch.
3Score a small batch of calls with Insight7. Use its self-serve call scoring and metered allowance on a few real support calls to see whether the output is useful. Book a demo for the voice and chat agents only if the scoring earns it.
Sources and checks
Below, by source, are the 9 figures we checked: 0 verified, 8 reported and 1 estimated.