builder, ai, markets, and general data nerd | https://xitter.cf/t.co/zs1Hb6u5e0
In the lab 🧪
Joined March 2025
- Tweets18.9K
- Following1.3K
- Followers1.6K
- Likes17.3K
Foliome now prices every goal in your household's plan and gives you everything you need to monitor and manage your wealth. Not just another dashbaord. Open source, runs on your machine.
It makes sense that AI removes the case-reading grind, but that grind was also where people got reps at judging what matters. If juniors skip it, the judgment he says stays human gets harder to build.
Open weights are nearly caught up.
Once Kimi K3.1 Pro's weights ship, the open vs closed gap on the Artificial Analysis Intelligence Index drops from 12 points to 3 (55 vs Claude Opus 5.5 at 58).
Most of the jump is agentic work. Terminal-Bench 4.0 went from 13% to 67%, ahead of Claude Sonnet 5.5 at 64%. Agent workloads burn many model calls per task, so that's where API bills pile up.
Companies running agents at volume could start pricing a self-hosted model against their closed-lab invoices, with only 3 index points separating the two.
Kimi K3.1 Pro launches at 4th on the Artificial Analysis Intelligence Index scoring 55, making large gains in agentic performance; it will be the leading open-weights model once its weights are published
@Kimi_Moonshot's new reasoning model gains 11 points over Kimi K3 (max, 44). It trails only Claude Opus 5.5 (max with fallback, 58), Claude Sonnet 5.5 (max with fallback, 56), and Muse Wonder Full (max, 56) on the Intelligence Index.
Once the weights for K3.1 Pro are released it will be the highest-scoring open weights model and the gap between the leading open weights and proprietary models on the Intelligence Index will narrow from 12 points to 3.
Key results:
➤ Large gains in agentic performance: Kimi K3.1 Pro scores 3rd on GDPval-AA with 1,846 Elo, up 312 Elo from Kimi K3, behind only Claude Opus 5.5 (1,866 Elo) and Muse Wonder Full (1,857 Elo). Kimi K3.1 Pro also achieves the highest score of any model on Terminal-Bench 4.0. K3.1 Pro scores 67%, up from 13% for Kimi K3 and ahead of Claude Sonnet 5.5 (64%). AutomationBench-AA, which tests agentic SaaS workflows, rises from 58% to 73%.
➤ Gains in other evals are more moderate: Humanity's Last Exam rises from 47% to 52%, CritPt from 23% to 25%, and SciCode from 59% to 61%. AA-Omniscience Index rises from 20 to 30, as the hallucination rate falls from 53% to 44% and accuracy rises from 48% to 51%.
➤ Cost per Task roughly equal to Kimi K3 at $1.98 per Task: K3.1 Pro uses 55K output tokens per Intelligence Index task, 14% more than Kimi K3 at 48K, while its first-party token prices are 33% lower.
Additional model details:
➤ Context window: 1M tokens.
➤ Size: 2.8T total, 104B active parameters
➤ Input modalities: Text and image
➤ Weights and license: Weights currently not published, Moonshot is expected to release the weights soon after launch
➤ Pricing: $2/$10/$0.20 per 1M input/output/cache hit tokens, 33% lower than Kimi K3 on all three rates.
Congratulations to the team at @Kimi_Moonshot on the release!
A defender with authority to move things inherits every weakness of whoever can talk to it. I'd limit it to freeze-and-alert and put money movement behind a hardware passkey, which won't hand credentials to a fake login page.
Open weights at 2.8T.
Only 104B parameters fire per token, about 3.7%. Compute per query follows the active count, but memory follows the full 2.8T, so self-hosting takes a rack of GPUs.
The 65% hallucination drop on finance modeling is a vendor number with no baseline stated. Once weights land Nov 4, funds and banks that can't send positions to an outside API can run their own eval on their own spreadsheets, inside their own walls.
Introducing Kimi K3.1 Pro: our most capable 3T-class open-weight model.
🔹 1M-token context, 2.8T total / 104B active params (MoE), Multimodal.
🔹 Full-project coding: games in Unreal, Unity & Godot.
🔹 Excels at parametric 3D modeling in Blender & CAD.
🔹 More reliable at finance work: Accurate modeling with hallucination down by 65%.
🔹 Matching the strongest proprietary models across coding, agentic and knowledge benchmarks.
Kimi K3.1 Pro and K3.1 Pro Highspeed (2–3× faster) are now live on Kimi.ai, Kimi Work, Kimi Code, and the Kimi API.
Open weights by Nov 4, 2026.
🔗 API: platform.kimi.ai
🔗 Tech blog: kimi.ai/blog/kimi-k3-1-pro
Hallucinations falling from 23.4% on K3 to 8.3% is a big deal for agent builders. Fewer babysitting loops, and once the weights land by Nov 4 you can self-host it.
Cyber sold off in March on the Mythos fears and now leads. If AI can find bugs that fast, someone still has to pay to patch them, so the same scare that hit the group becomes the bull case.
Sheesh. 65.7 vs 66.4 against Opus 5.5 on Terminal-Bench 4.0 is basically a tie on agentic coding, and once the weights drop Nov 4 anyone can run it.
Both sides get the same unflagged model, so the edge goes to whoever already has a working patch process. Attackers need one good exploit chain, defenders need the boring part done: inventory, patching, rollout.
Rough for the doom crowd. Amodei's window was one to five years for half of entry-level white-collar jobs though, and that hit shows up as junior roles that never get posted well before it shows up in layoffs.
Feels like the cars-or-guns call gets made the day weights go public, because abliteration strips the safeguards right after release. The open question is patch speed: defenders can run the same bug-chaining on their own code first.
K3.1 Pro lands 0.7 points behind Opus 5.5 on Terminal-Bench 4.0 and ahead of GPT-6 Astra, with weights public Nov 4. The closed labs' lead is down to a few weeks of release timing.
Biotech is already moving. Anthropic paid roughly $400M for Coefficient Bio and is building wet-lab infrastructure, so its unreleased model runs experiments on its own bench. A startup renting the public version is months behind the house team on every run.
Heineken can't out-market a drug that turns off the craving, so it's following the people who quit. Its non-alcoholic line grew 12% in Q3 2025.
Jassy's "deploy in minutes" point is already playing out. Dutch and Finnish government sites went dark this week waiting on a PHP patch that was already published, and Cloudflare was seeing early attacks in the gap.
Finding bugs is cheap now, but patching is still slow. Two PHP flaws were disclosed and within days the Netherlands and Finland were taking government sites offline because they couldn't patch fast enough.
Stanford's AI Index has Southeast Asia as the most trusting of its own governments to regulate AI, too. Optimism comes easier when you think someone competent is steering.
Build, then tap through it in the simulator, then fix what broke. That closes the loop most agents leave open. Since it works off screenshots, every step also leaves a debugging trail you can replay.
Big hire. Greenwald knows how government buys, and Jordan's Siraj (1.3M users across 5,800 schools) gives him a real deployment to show agencies from day one.