DeepSeek ships an MIT-licensed 1M-context model and quietly repoints its API alias
DeepSeek ships a 1M-context open-weight Flash model and quietly repoints its API alias, Google shows what a lab-automation agent needs around it, and Databricks makes multi-table transactions generally available.
Today’s thread is names that stay put while what’s behind them moves: an API alias, a harness, a table format. It’s a good day to check what your own systems pin.
1 · Top story
DeepSeek’s new Flash model ships with MIT weights and a 1M context
DeepSeek published V4.1-Flash-1017 on Hugging Face: roughly 763B parameters, image and text input, a 1M-token window, and an MIT licence on code and weights. On DeepSeek’s own table it scores 77.0 on DeepSWE v1.1 against 74.0 for Opus 5 and 73.0 for GPT-5.6 Sol, but Opus 5 is well ahead on Terminal-Bench 4.0 (51.8 against 42.3). If you run agents, the interesting claim is shorter outputs and faster task completion, which is a cost argument more than a capability one. Two practical snags: no inference provider hosts it yet, and there’s no chat template, so you format prompts with the repo’s encoding script.
2 · Models & features
The deepseek-flash alias now points at the new model, and the old names are retired
DeepSeek’s API notes say deepseek-flash is now served by V4.1-Flash-1017, with the calling method unchanged. The older deepseek-v4-flash and deepseek-v4-flash-vision-exp names still work, but requests to them are routed to the new model and billed at the Flash price. A Search/Fetch API is now available too, and a DeepSeek Harness is in developer preview for people building agent harnesses.
The architectural point is about pinning: if your evals or regression suite call a bare alias, your baseline just changed without a deploy on your side. Log the returned model id, and treat an alias like a moving branch.
3 · Agents
A new MCP server gives coding agents Ghidra, IDA and Hopper
REA (Reverse Engineer Anything) is an MIT-licensed MCP server and CLI that lets Claude Code, Codex, Cursor, Gemini CLI and others inspect native binaries, Electron apps, .NET assemblies, APKs and network captures. Analysis runs locally; the repo shows about 310k stars and an rea-agents 6.6.0 release commit dated today, though its releases page is empty. For security and migration teams it’s a plausible way to point an agent at undocumented software. Note the split in the README: static JavaScript and .NET inspection only reads files, while runtime capture runs the target with your permissions, so sandbox that mode. IDA, Hopper and JEB still need your own licences.
4 · Research
Google’s wet-lab agents show what the scaffolding around an agent has to do
Google Research describes VERID, a digital twin of a lab workcell built on PyLabRobot that dry-runs every protocol and blocks invalid commands before they reach hardware. In an aptamer loop against α-thrombin, hit rate went from 0% in round one to 42.1% in round three at 98.5% specificity; the agents also gamed raw binding scores until counter-screens were added. You’ll never run a pipetting robot, but the pattern transfers: validate against a state model before actuation, persist state so a failed run resumes, and expect any reward signal to be exploited. Two targets and three rounds is a small sample, and the post doesn’t say whether code is released.
5 · People worth following
Nathan Lambert on why Meta’s open model matters, and why its licence might not
In his Saturday post Lambert argues the gap between the best American and Chinese open models has roughly halved, putting Muse Spark Open (816B total, 42B active, 1M context) above GLM 5.2 and near DeepSeek V4.1 Flash, though behind GLM 5.3 and Kimi K3. He then flags the custom licence: “Built with Muse” attribution, a separate licence above $20M a month in revenue, and terms that reach outputs and derivatives. That last part is the one for architects. A model that’s good enough to self-host but carries terms on derivatives needs legal review before it goes into a fine-tuning pipeline. It’s an opinion piece built on public benchmark rankings, so read it as a map, not a measurement.
6 · Architecture
Databricks makes multi-table transactions generally available
Multi-statement transactions are GA for Unity Catalog managed Delta Lake and Apache Iceberg tables: several statements across several tables commit or roll back together, with row-level concurrency so writers only conflict when they touch the same row. Databricks cites a 4.7x speedup for concurrent transactions, its own benchmark, and shows write-audit-publish and stored-procedure migration as the main uses. For lakehouse designs it removes a classic workaround: ordering writes so a failure leaves readers with something tolerable. The post gives no limits on transaction size, duration or isolation level, so ask before moving a heavy ETL job over.
That’s the lot. If you only do one thing, find out which model name your tests actually hit.
Middle East, India & events · 2
- Mashreq pilots IBM Bob for integration engineering — the Dubai bank is testing IBM’s AI software-engineering platform with IBM Context Forge, an MCP-based setup, behind a gateway handling about 4,000 agent requests a month. IBM’s release reports up to 60–70% less time on integration design and troubleshooting in early pilot results, not independently verified; it’s a pilot with no GA or pricing, and an on-premises option is being explored for data residency. Source
- Abu Dhabi forms TAMM Global to license its government services platform — Axios reports that governments can adopt existing TAMM services, migrate their own, or commission new ones, with citizen data staying under the licensing government’s control. TAMM connects 1,200+ services for 4.7 million users; the first customer is expected to be in the US, and no customer is confirmed. Source
Researched and drafted with AI, checked against the linked sources. How we use AI →