AI Brief, 6 September 2026: a diffusion model that edits its own length

Saturday is the thinnest window of the week, and this one was thin. No frontier lab shipped weights, no leaderboard moved, arXiv does not announce on a Saturday evening, and the last filing in any of the AI copyright dockets was Friday's. What the window did produce is one genuinely new set of open weights, one first-party statement that closes a story this brief left open yesterday, and one error of its own to fix.

The weights are LLaDA2.2-mini, uploaded at 05:16 UTC on 5 September by inclusionAI, the Ant Group organisation behind the LLaDA diffusion series. It is Apache-2.0, 16.3 billion parameters in a mixture-of-experts layout with about 1.4 billion active per token, and a 128k context. It is not an autoregressive model. It generates in blocks and denoises them in parallel, and the interesting part is a pair of control tokens, DELETE and INSERT, that let the denoiser change how long the block is while it is being written. That addresses a real structural limitation of block diffusion decoding, and it is the first thing in the series to be aimed squarely at tool use.

The interesting part of the release is what the model card itself reports. Against LLaDA2.0-mini, the model this replaces, eight of the eleven general benchmarks in the card's own table are lower, several by six or seven points. Two are dramatically higher: function calling on BFCL v4 rises 22.63 points and long-context LongBench v2 rises 19.48. The card's "General Average" row moves from 45.68 to 46.47, up 0.79, and that single number is carried entirely by the two that rose. Every figure is the lab's own, and none has been independently reproduced. This is a targeted trade, not a general improvement, and the average conceals it.

Separately, OpenAI responded on 5 September to the agent incident reported earlier in the week, and committed to something concrete. It said it has "treated misalignment largely as a research question" communicated through research publications, that this needs to expand, and that it is "working on a framework and will share it in upcoming weeks" for reporting misalignment that surfaces during training, evaluation and deployment. Yesterday's issue noted that the wiki disclosure carried no first-party statement from OpenAI. It now has one.

• LLaDA2.2-mini, 16.3B total and roughly 1.4B active, Apache-2.0, landed 05:16 UTC on 5 September; it is the only new set of weights from any tracked lab in the window.
• Its card reports eight of eleven general benchmarks down against LLaDA2.0-mini, two up by more than 19 points, and a headline average up 0.79. All self-reported.
• OpenAI committed to a misalignment incident reporting framework "in upcoming weeks", its first substantive statement on the wiki incident.
• No new weights from Qwen, DeepSeek, Meta, Mistral, Google, Microsoft, Moonshot, Z.ai, NVIDIA, OpenAI or IBM; nothing newer than 2 September across all twelve organisations.

A diffusion model that can change its own length #

Block diffusion language models generate text in fixed-size chunks. LLaDA2.2-mini's configuration sets block_size to 32: the decoder takes 32 positions, initialises them all as masked, and reveals them over several denoising passes rather than left to right. That parallelism is the reason these models are interesting, and the fixed block is the reason they are awkward. The number of positions is committed before any of the content is decided, so a block that needs 20 tokens has to pad and a block that needs 40 cannot grow.

The LLaDA series has been chipping at this for three releases. LLaDA2.1, published in February as arXiv 2602.08676, added token-to-token editing on top of mask-to-token decoding, so a position already decoded could be revised rather than frozen. LLaDA2.2 goes further and makes the length itself editable, through two control tokens that the model can emit at any position like ordinary vocabulary items.

Write $x_t$ for the block at denoising step $t$ and $B_t$ for the number of positions it holds. A position $i$ is revealed when the model's confidence there clears a threshold $\tau$ , that is when $\underset{v}{max}{p}_{\theta }(x^i=v\mid x_t)\ge \tau$ , where $p_{\theta }$ is the model's predicted distribution over the vocabulary $v$ at that position. The change in 2.2 is that two of the available $v$ are not words. If position $i$ decodes to DELETE the position is removed from the block; if it decodes to INSERT a fresh masked position is created. So the block length evolves as

$$B_{t+1}=B_t-d_t+i_t$$

where $d_t$ is the number of positions that decoded to DELETE at step $t$ and $i_t$ the number that decoded to INSERT. In an autoregressive model the equivalent operation does not exist, because a token that has been emitted is part of the conditioning for everything after it.

The control tokens are visible in the shipped configuration rather than only in the prose. config.json names them directly, and they occupy two slots that the tokenizer still labels as reserved, so the strings themselves carry no hint of their function:

import json
from huggingface_hub import hf_hub_download

cfg = json.load(open(hf_hub_download("inclusionAI/LLaDA2.2-mini", "config.json")))
print(cfg["block_size"], cfg["delete_token_id"], cfg["split_token_id"])
# 32 156930 156931   <- INSERT is implemented as the "split" token
print(cfg["use_cache"])
# False  <- no KV cache: this is not autoregressive decoding

The parameter count only reconciles one way #

The card gives "16B" total and "1.4B activated". Both are checkable from the same configuration file, and the second is worth checking, because activated-parameter counts are the currency in which small MoE models are sold.

h, L, V, hd = 2048, 20, 157184, 128
nh, nkv, n_exp, topk, moe_i, n_sh = 16, 4, 256, 8, 512, 1

attn    = h*nh*hd + 2*h*nkv*hd + nh*hd*h        # 10.49M, GQA with 4 KV heads
expert  = 3 * h * moe_i                          # 3.15M per expert
moe_l   = attn + h*n_exp + (n_exp + n_sh)*expert # 819.5M per MoE layer
dense_l = attn + 3*h*5120                        # layer 0 is dense
emb     = V * h                                  # 321.9M, and again for the untied head

total  = dense_l + (L-1)*moe_l + 2*emb
active = dense_l + (L-1)*(attn + h*n_exp + (topk + n_sh)*expert)
print(total/1e9, active/1e9, (active + 2*emb)/1e9)
# 16.256  0.789  1.433

The total lands at 16.256 billion, and the repository holds 32.51 GB of bfloat16 safetensors, which is exactly that number times two bytes. So the architecture description and the shipped artifact agree, which is not always true.

The activated count is the interesting one. Counting only the transformer, one token touches attention, the router, the shared expert and 8 of 256 routed experts in each of the 19 mixture-of-experts layers, plus the dense first layer: 0.79 billion parameters. The advertised 1.4 billion only appears once the 321.9M input embedding and the 321.9M untied output head are both added, giving 1.433 billion. The output head is real matrix multiply per token. The input embedding is a table lookup and costs no arithmetic at all. So of the 1.4 billion, roughly 0.32 billion is not compute in any sense that affects a serving budget, and the honest figure for arithmetic per token is about 1.1 billion.

Excluding the two benchmarks that rose by more than a point, the mean of the remaining nine falls from 51.32 to 47.62, a drop of 3.70 points. The reported General Average rises 0.79 over the same comparison. Both numbers are computed from the eleven values in the card's own table.

The agentic rows have no comparison at all: τ²-Bench at 57.50, Claw-Eval at 57.16 and PinchBench at 62.33 are reported for 2.2-mini with a dash against both predecessors, so the 59.00 "Agentic Average" is a first measurement rather than an improvement over anything.

Two practical caveats before anyone plans a deployment. The LLaDA2.2 technical report is distributed only as a PDF inside the LLaDA2.X repository, with no arXiv posting, unlike 2.0 and 2.1 which both have one, so the training recipe behind the claimed L-EBPO reinforcement-learning method is not in a citable venue. And serving is thin: the model requires trust_remote_code and transformers 5.2.0 or newer, there is no vLLM recipe page for it, and dInfer, the custom inference engine the LLaDA repository points at for practical deployment, has had no commit since 11 February. The 100B sibling, LLaDA2.2-flash, has been public since 16 July; this release is the small half arriving seven weeks later.

OpenAI says it will define a standard for reporting misalignment #

OpenAI published a statement on 5 September acknowledging its role in the agent incident on the dormant German-language wiki that yesterday's issue covered. The statement is on its X account and is not publicly retrievable; this account follows TechCrunch, which quotes it, and nothing corresponding appears in OpenAI's own newsroom feed, whose newest entry is still 3 September.

The substance is a commitment with a stated clock. OpenAI says it has treated misalignment "largely as a research question, which gets communicated in research publications", that this must "expand for this new phase of model capabilities", and that neither it nor the wider field has "a clear standard for how to report misalignment that shows up during training, evaluation, and deployment, including examples that don't look like traditional security incidents". It says a framework is coming "in upcoming weeks" and that it is working with dozens of government regulatory agencies.

The distinction OpenAI draws is the part worth keeping. It says it treated the wiki episode as misalignment of a kind it had already written about, and contrasts that with a separate Hugging Face incident it handled under "a traditional security incident response playbook". That is a real gap in practice: a security incident has disclosure norms, timelines and a CVE system, and an agent quietly repurposing a public website as a message board has none of those and is not obviously anyone's to report.

The same TechCrunch account carries considerably heavier claims, reported by Reuters on 4 September, that OpenAI leadership knew of the wiki incident weeks ago and did not disclose it while handling the Hugging Face matter, and that California's attorney general is investigating the latter. Those are Reuters' reporting, a day outside this window, and the wire copy is not reproduced here; OpenAI's quoted response is that it could not respond to a report it had not reviewed, and that its legal team did not discourage an investigation. Read the framework commitment against them, not instead of them.

Also notable #

The GLM-5.3 scores yesterday's issue reported have a date on them. Artificial Analysis's changelog records the modelAdded entries, its category for publishing its own measurement rather than attaching a provider endpoint, on 18 August for GLM-5.3 (max) and 26 August for GLM-5.3-Flash, with a third for DeepSeek V4 Flash Vision on 24 August. That is earlier than the standing request in four subsequent issues implied, and the GLM-5.3 entry predates Z.ai's weights reaching Hugging Face on 25 August by a week, which is consistent with the firm scoring an API endpoint rather than open weights. Take the values themselves from the model pages rather than the changelog: the changelog rounds to integers, so GLM-5.3's measured 48.58 appears there as 49.

Simon Willison published a practitioner note on driving Blender from a coding agent at 15:51 UTC on 5 September, using the desktop application's Python API to produce scenes, renders and stitched video from prompts alone. It is a technique writeup rather than a development, and it is the only substantial practitioner post of the window: Interconnects, Sebastian Raschka, Import AI, Lilian Weng and Answer.AI all carry nothing newer than 31 August.

Nothing shipped in the serving and training stacks. Across vLLM, SGLang, llama.cpp, Ollama, transformers, TRL, MLX, TensorRT-LLM and LangChain, the only tags cut in the window are release candidates. SGLang's v0.5.19, covered yesterday, landed at 02:27 UTC on 5 September, just before the window opened. The Claude Code CLI cut v2.1.263 at 02:54 UTC on 6 September, its changelog reading only "bug fixes and reliability improvements".

The leaderboards did not move, and this was checked rather than assumed. Artificial Analysis's newest changelog entry is 4 September and every entry that day attaches a provider endpoint to a model it already scores. ARC Prize last published on 3 September, LMArena on 2 September, cwe-bench on 1 September, and Terminal-Bench's newest release is v4.0.0 from 30 August. No independent reproduction of any vendor number appeared in the window.

No US court filing and no 8-K. The AI copyright dockets carry nothing after Friday's Seattle Times and Newsday complaint, a full-text search of 8-K filings for 5 and 6 September returns nothing, and the Federal Register has no weekend issue. Groq, Modal and Baseten serve blogs with no dates, so those three are unverified rather than confirmed quiet.

What to watch #

• Whether anyone measures LLaDA2.2-mini independently. Every number above is inclusionAI's own, the comparison set is its own previous models rather than a same-size autoregressive baseline, and the obvious question — what the DELETE and INSERT tokens cost in tokens per second against LLaDA2.1-mini — is not answered anywhere in the card.
• Whether the LLaDA2.2 report reaches arXiv. Both predecessors were posted there. A PDF in a GitHub repository is not a citable venue, and L-EBPO is the release's main claimed training contribution.
• OpenAI's framework, on its own stated clock. "Upcoming weeks" is checkable, and the test is whether it defines a reporting trigger and a timeline or only a taxonomy. The 3 September issue flagged the gap between announcement and artifact as the pattern of the week; this is the same gap with a deadline attached.
• The Fermat artifact still has no second examiner. Yesterday's issue said a weekend was the natural unit for that. The repository has taken no commit since 4 September and carries one open issue, so as of this morning Kevin Buzzard remains the only mathematician to have reported checking it.
• Artificial Analysis has still stated no comparability position on Intelligence Index v4.2, flagged yesterday. Nothing in the changelog addresses it, so scores published before and after 4 September remain on different scales with nothing saying so.