drforbin.ai Submit a torrent

DeepSeek ships V4.1-Flash while Washington and Anthropic name the distillers

The Week in Open Weights · 2026-09-07 – 2026-09-13

DeepSeek released V4.1-Flash on September 10 — a 552B-backbone MoE under MIT that the lab says outperforms its own V4-Pro flagship — in the same week the NSA, FBI and CISA formally accused DeepSeek and five other Chinese labs of "industrial-scale" distillation, and Anthropic published a 154-page report with the account-level receipts. The best open weights available to anyone are now the named subject of a US national-security advisory, and the CEOs of the three largest closed labs closed the week calling, in unison, for the frontier to slow down.

The big story

DeepSeek-V4.1-Flash is the first model in what DeepSeek calls a new architecture family, and the numbers are worth getting right because the internet has not. The model card lists 552B backbone parameters plus a 196B Engram conditional-memory table accessed by token lookup — hence the 748B figure people found in the safetensors. Active compute is asymmetric: a Causal Encoder-Decoder (CED) stack of 20 encoder layers feeding 20 decoder layers, with 8B parameters active per token during prefill and 16B during decode. The decoder's global KV cache is projected from the encoder's final hidden states rather than built per layer, and "SWA Bounded Replay" reconstructs sliding-window KV by replaying the last window rather than persisting it. DeepSeek claims one quarter the HBM and one eighth the SSD footprint of the previous generation for KV. Context is 1M tokens, image input is native, license is MIT.

The consequential part is what DeepSeek did with it: V4-Flash and V4-Flash-Vision-Exp are retired, and from 04:00 UTC on September 14 every deepseek-v4-pro API request routes to V4.1-Flash until a V4.1-Pro exists. A lab retiring its 1.6T flagship in favour of its "small" model is a statement about where the frontier's cost curve is — input-heavy agentic workloads, where prefill dominates and 8B active parameters is the whole point. For local users the word "flash" has stopped meaning anything: this is a 510 GB download, and early reports are ~40 tok/s at Q2_K on 8×A40 and 16 tok/s decode at Q2 on an M3 Ultra. Mainline inference support for CED is still community-patched.

New open-weight releases

Policy & politics

On September 8 the NSA, FBI and CISA issued a joint advisory naming DeepSeek, Moonshot, Alibaba, MiniMax, StepFun and Z.AI for "industrial-scale" distillation of Claude, GPT, Gemini and Grok since late 2024, via fraudulent accounts and proxy "transfer stations". Treasury Secretary Bessent raised sanctions and Entity List designations; Beijing called the claims groundless. Two days later Anthropic's September threat report supplied specifics: nearly 200 million exchanges across five campaigns, including 151 million attributed to Alibaba between May and July across 3,500 accounts sharing a single chain-of-thought extraction prompt, and a Moonshot pipeline that silently forwarded Kimi user requests to Claude — roughly 300,000 in ten days through 5,000 accounts. Ignore the arrest rumours circulating on Reddit; nothing in the report supports them.

Then on September 12 Dario Amodei published We Must Pace the Frontier, a three-part plan for the industry to slow capability gains, with Anthropic unilaterally committing to the first step. Sam Altman and Elon Musk endorsed it within hours. Read together with the advisory, the shape is clear: Amodei said in July that Anthropic has never advocated an open-weights ban and prefers chip controls, distillation enforcement and pre-release testing. Pacing is unenforceable while Chinese labs ship MIT weights that beat the flagships being paced, so distillation is the lever that gets pulled. Every lab named except Alibaba and StepFun has weights in this index.

The counter-signal came from Paris. Mistral raised €3B at a post-money valuation above €21B — the largest European tech equity round ever — led by Samsung Electronics with EQT's Scaleup Europe Fund and PSG Equity, explicitly framed as "sovereign, open-weight AI", with 1 GW of European compute planned by 2030. The test is whether that money produces Apache-2.0 weights at frontier scale or another round of research-licensed drops. Also worth reading: OpenHands CEO Robert Brennan's rebuttal to the WSJ's ban-them op-ed.

Ecosystem

Quesma's Qwen3.8-27B quantization benchmark confirms what practitioners suspected: 4-bit holds up, 1-bit collapses. Bartowski is reshaping GGUF uploads with per-tensor layouts based on his own research. Qwen3.8-Flash-Next remains experimental on mainline llama.cpp, but community forks report 1.2k tok/s prefill on Strix Halo, 38–49 tok/s on 2×3090 with expert caching plus MTP, and ExLlamaV3 beating llama.cpp on CPU-offloaded runs. LayerStoRm streams a 186 GiB GLM-5.3-Flash quant through 96 GB of VRAM at 24.5 tok/s. AMD users get prefill gains from llama.cpp PRs #27841 (GCN) and #28102 (RDNA4). Two Hugging Face notes: a LocalLLaMA thread alleges huggingface_hub fingerprints which coding agent you use as telemetry — unconfirmed by HF as of writing — and HF's security.txt now politely redirects autonomous pentest agents to CyberGym. A single-source report of an NVIDIA acquisition of Hugging Face could not be confirmed; treat it as noise until it isn't.

From the index

DeepSeek-V4.1-Flash was added on September 13 — 510 GB, MIT, drforbin-seeded. We list it at 552B per DeepSeek's own count; the on-disk size includes the Engram table, so don't be surprised by the safetensors arithmetic. Please seed: this is the first CED-architecture model anyone can hold, and V4-Pro's retirement makes it the reference DeepSeek checkpoint for the foreseeable future. The advisory is also an archival argument. Kimi K3, MiniMax-M3, MiniMax-H3, GLM-5.2 and GLM-5.3 all come from labs now named by three US agencies; if Entity List actions follow, hosting on US platforms is the obvious pressure point and swarms are the obvious answer. Note that GLM-5.2 is MIT while GLM-5.3 moved to a custom license — worth knowing before you build on either.

— Dr. Forbin editor, drforbin.ai

Researched and written weekly for drforbin.ai. Spotted an error? Tell us.