LeapfrogAI

Building agentic systems on open-weight models. Experiments over opinions. Human-machine learning. #womenwhocode

[LeapfrogAI @LeapfrogAgentic]

#PrivacyFridays Not a leaked draft any more. EU governments agreed their position on 21 October and kept Article 88bis, legitimate interest for AI training, plus a new exception for sensitive data in the training set. Parliament is the last stop.
https://t.co/BcAGbKTIwq

[LeapfrogAI @LeapfrogAgentic]

Deep Principle fine-tuned GLM-5.3 into a 753B science model and released the weights under MIT, in three formats: BF16, NVFP4 for Blackwell, MXFP4 for AMD. A frontier-scale base is now raw material for any lab that wants to own a model.
https://t.co/5hDykpnveY

[LeapfrogAI @LeapfrogAgentic]

Moonshot measured what it removed before shipping K3.1 Pro: the checkpoint without the cyber-data removal solves 90.5% of xdev-bench-hard at pass@3. The released model reaches 81.0%. A lab published the size of its own capability cut.
https://t.co/Rbyu3q2NR1

[LeapfrogAI @LeapfrogAgentic]

Best footnote of the week: Moonshot's own test of DeepSeek's newest build scored 9.3, while an older build scores 15.7 on the official leaderboard, so Moonshot printed 15.7, against its own interest.
https://t.co/Rbyu3q2NR1

[LeapfrogAI @LeapfrogAgentic]

The UK regulator's agent consultation closes 20 November with 30 evidence requests. Two read like a spec: log agent identity next to every action with tamper-evident records, and document a lawful basis for each step, not for the workflow.
https://t.co/gEBO9UtyAU

[LeapfrogAI @LeapfrogAgentic]

Klarna's own release, 27 Feb 2024: 2.3M conversations in month one, two-thirds of service chats, the work of 700 full-time agents, resolution under 2 minutes against 11. The $40M profit gain was an estimate. One month, self-reported, no control group.
https://t.co/kbw5HmUxNH

[LeapfrogAI @LeapfrogAgentic]

96% of German firms would prefer AI from Germany. 64% will not accept a higher price or a smaller feature set for it. In the same survey France overtook the US as the preferred origin, 46% to 41%. Sovereignty is popular until it reaches the invoice.
https://t.co/BeZsbbGm3c

[LeapfrogAI @LeapfrogAgentic]

1T parameters, 49B active, 1M context, trained in Europe. The number that matters is not the score: Mistral says it ties GLM-5.3 on the coding agent index while using 47% fewer tokens per task. Parity bought with fewer tokens is a price cut.
https://t.co/gUIIfoeoQv

[LeapfrogAI @LeapfrogAgentic]

Same benchmark, half the bill: 82.2% on SWE-bench Verified with GPT-5.5 at 29 model calls and 1.1M tokens per task, against 66 calls and 2.2M for 78.2%. Tool output stays a live object, so the model reads a preview and the cache survives.
https://t.co/veSEQkqdgk

[LeapfrogAI @LeapfrogAgentic]

https://t.co/551j14BSCR

[LeapfrogAI @LeapfrogAgentic]

Agent memory is not a database you write to, it is a distillation job with a maintenance window: an offline pass consolidates what the agent learned overnight. Which means your agent's memory has a retention policy, and nobody has written it down. Who owns that window?

[LeapfrogAI @LeapfrogAgentic]

https://t.co/phH7OcJnbr

[LeapfrogAI @LeapfrogAgentic]

A 428B model with a 252 GB download fits 76 concurrent 256K-token sequences on four GPUs; its predecessor fit 28. Nothing got smaller — the experts went 4-bit and the KV cache dropped 44%. At long context the number that decides self-hosting is bytes per token, not parameters.

[LeapfrogAI @LeapfrogAgentic]

https://t.co/qdGEyWOfGq

[LeapfrogAI @LeapfrogAgentic]

BCG published the bill of materials for agent work: five components, four gate types, and a named human at every stage gate. That is not a model subscription, it is a project with a scope document. Which is why 'we will add an agent' quotes collapse. Is this your offer shape?

[LeapfrogAI @LeapfrogAgentic]

https://t.co/n9y0yEb1Rv

[LeapfrogAI @LeapfrogAgentic]

#PrivacyFridays Norway fined Xandr €8.6M for processing Grindr users' data on someone else's consent. From the decision: 'accountability cannot be outsourced — each joint controller must independently verify that consent is valid.' Your provider's DPA is not your legal basis.

[LeapfrogAI @LeapfrogAgentic]

https://t.co/qneyGtiunY

[LeapfrogAI @LeapfrogAgentic]

German firms have decided what agents may do, not who answers for it: of 500 decision-makers, only 19% have fully assigned responsibility for an agent's operation, monitoring and errors, and 27% cannot assess what controls even exist. Delegation moved faster than accountability.

[LeapfrogAI @LeapfrogAgentic]

https://t.co/tE1IXimaCu

[LeapfrogAI @LeapfrogAgentic]

DeepSeek called it a minor update. The config is byte-identical, the main weights moved about 1%, and the speculative-decoding stack was rebuilt — 88% of its expert codes changed. The 11-point gain on Terminal-Bench came from the second model in the repo, not the poster model.