AI Brief, 22 September 2026: a trillion-parameter model that ships in four bits

Xiaomi published MiMo-V2.6 on Monday 21 September under an MIT licence. The flagship checkpoint, MiMo-V2.6-Pro-RL, is a sparse mixture-of-experts model the card describes as 1.02 trillion total parameters with 42 billion activated per token, taking text, images, video and audio, with a context limit of 1,048,576 tokens. The weights began uploading at 19:34 UTC and the model card went up at 20:12. A smaller sibling, MiMo-V2.6-Flash-RL, landed alongside it, and a 9-billion-parameter distillation onto a Qwen base followed two hours later.

The number that matters is not the trillion. It is 534 gigabytes. That is the entire size of the flagship checkpoint, because Xiaomi did not release a bf16 model and then quantise it: the routed expert weights, which are 97.7% of the parameter count, are stored in 4-bit MXFP4 in the only checkpoint there is. Held in bf16 the same model would be 1.86 TiB. Both figures can be checked against the arithmetic in the configuration file, and they agree exactly, which is rarer than it should be. The practical consequence is that a trillion-parameter model now has a published serving recipe for a single eight-GPU node, and the company's own instructions use one.

The capability claim is harder to assess than the storage one. Every figure for MiMo itself is Xiaomi's own, three of the seventeen benchmark rows are benchmarks Xiaomi wrote, and the predecessor scores in the comparison column include a 0.0 and a 1.5, which usually describes a harness failing rather than a model failing. The independent reading arrived separately and was kinder than that suggests: Artificial Analysis measured the Pro checkpoint at 46.3 on its Intelligence Index, and on the one benchmark where its numbers and Xiaomi's can be laid side by side they agree to a tenth of a point. xAI released Grok 4.7 the same day, proprietary and priced at $2 and $6 per million input and output tokens, and the same firm measured it at 46.4. One of those two models can be downloaded.

Also on Monday: the 1.21 terabytes of ungated Step 5 Preview weights this account flagged on the 20th are no longer public, resolving that question in the direction the absence of a licence suggested.

• MiMo-V2.6-Pro-RL: 1.02T total, 42B activated, MIT, 1M context, omnimodal, 534 GiB on disk.
• Every one of its 1,000,190,509,056 routed-expert parameters is stored in 4 bits; nothing else is.
• Hugging Face's own parameter counter reads 524 billion for it, because it counts packed bytes.
• Grok 4.7 (xAI, proprietary, $2/$6 per M tokens) scored 46.4 on Artificial Analysis's index against MiMo-V2.6-Pro's 46.3. Both measurements are independent; both models' own claims are not.
• StepFun's Step 5 Preview weights have gone behind an access wall and out of the public listing.
• The UN's science panel published its first thematic brief, on agent misalignment, built around the OpenAI-Hugging Face incident covered here on 17 September.

A trillion parameters, engineered to fit #

Start with the storage claim, because it is checkable and because checking it turns up something the release does not mention.

config.json for the Pro checkpoint declares 70 layers, hidden size 6,144, 384 routed experts with 8 active per token, no shared experts, and an expert intermediate size of 2,048. The first block is dense; the remaining 69 carry MoE feed-forward networks. Each expert holds three matrices — gate, up and down — so the routed-expert parameter count is fixed by those five numbers:

$$N_{\text{experts}}=E\cdot 3\cdot d_{\text{model}}\cdot d_{\text{ff}}\cdot L_{\text{moe}}$$

with $E=384$ experts, $d_{\text{model}}=6144$ the hidden size, $d_{\text{ff}}=2048$ the expert intermediate size, and $L_{\text{moe}}=69$ the number of MoE layers. That is 384 × 3 × 6144 × 2048 × 69 = 1,000,190,509,056 parameters.

Now ask the repository what it actually contains. The checkpoint's tensor metadata reports 500,095,254,528 values of type U8, alongside 10.6 billion in BF16 and 13.4 billion in FP8. Double the U8 figure and you get 1,000,190,509,056 — the routed-expert count, to the byte. Every expert weight is a 4-bit value packed two to a byte, and every other weight in the model, including all attention projections, both embedding matrices, the vision tower and the audio encoders, is not.

import json, urllib.request

url = "https://huggingface.co/api/models/XiaomiMiMo/MiMo-V2.6-Pro-RL"
p = json.load(urllib.request.urlopen(url))["safetensors"]["parameters"]

p["U8"]                                        # 500_095_254_528  packed pairs, not parameters
p["U8"] * 2 + p["BF16"] + p["F8_E4M3"]         # 1_024_216_576_896  the real count

This is why the model's Hugging Face page advertises roughly 524 billion parameters: the counter treats each packed byte as one value, and halves the expert stack. The card's 1.02T is the honest number and the site's own headline figure is the misleading one.

The quantization_config block says how, and the two fields that matter are easy to miss:

"quantization_config": {
  "quant_method": "fp8",
  "store_dtype": "mxfp4",
  "mxfp4_block_size": 32,
  "weight_block_size": [128, 128]
}

MXFP4 gives each weight a 4-bit value and shares one 8-bit exponent scale across a block of 32, so the amortised cost is 4 + 8/32 = 4.25 bits per weight. Measured across the whole checkpoint the figure is 4.479 bits per parameter — 573,457,361,664 bytes for 1.024 trillion values — with the excess accounted for by the bf16 and fp8 tensors that were never packed.

The context length is a memory decision too #

The same discipline shows up in the attention stack. The config's hybrid_layer_pattern marks 10 of the 70 layers as global attention and the other 60 as sliding-window, and sliding_window is 128 tokens. That is a very short window: 60 of 70 layers never see more than 128 tokens of history directly. The card's table agrees, giving 70 layers as 60 SWA and 10 GA, with global attention in the first block and roughly one global layer in every seven thereafter.

That ratio decides whether a million-token context is affordable. Each layer caches keys and values for 8 key-value heads, with a key head dimension of 192 and a value head dimension of 128, so in bf16 one token costs $n_{kv}(d_k+d_v)\cdot b$ = 8 × (192 + 128) × 2 = 5,120 bytes per layer, where $n_{kv}$ is the number of key-value heads, $d_k$ and $d_v$ the key and value head dimensions, and $b=2$ bytes for bf16. A sliding-window layer stops accumulating at 128 tokens; a global layer does not. At the full 1,048,576-token context:

• 10 global layers × 1,048,576 tokens × 5,120 bytes = 50.0 GiB
• 60 sliding-window layers × 128 tokens × 5,120 bytes = 37.5 MiB

If all 70 layers were global, the same context would cost 350 GiB of cache. The hybrid pattern cuts it by exactly the ratio of the layer counts, 70 to 10, and the saving is what makes the advertised context length something other than a specification-sheet number.

Put the two halves together and the design target is obvious: 534.07 GiB of weights plus 50.0 GiB of cache at maximum context is 584 GiB, against the 596 GiB of usable memory on a node of eight 80 GB accelerators. Xiaomi's published vLLM command runs at --tensor-parallel-size 8 with --gpu-memory-utilization 0.95, which is the fit those numbers imply and not a comfortable one.

The caveats are the ones the release invites. The table carries no harness, temperature or task-count footnotes, and does not say which cells are Xiaomi's runs and which are quoted from elsewhere. Three of its rows — MiMo Code Bench, MiMo Cyber Bench and MiMo VisualCoding — are benchmarks Xiaomi authored, and no other lab's model can be checked against them. The MiMo-V2.5-Pro comparison column records 1.5 on Terminal Bench 4.0 and 0.0 on MiMo Cyber Bench, figures that describe a model failing to run rather than a model performing badly, which flatters every improvement in the same row. A technical report ships in the repository. Its claims about the training recipe, an asynchronous GRPO run at 1,568 prompts by 16 rollouts per step with an agentic grader ranking rollouts within each group, are not independently checkable from the artifact.

The same model, the same benchmark, two numbers #

xAI's Grok 4.7 went out on 21 September, priced identically to Grok 4.6 at $2 and $6 per million input and output tokens, and pitched as "twice as fast, at half the price of comparable models". Artificial Analysis put it sixth on its Intelligence Index at 46.4, behind Claude Fable 5.1 at 53.4, GPT-6 Astra at 52.7 and Claude Opus 5 at 50.8, and a tenth of a point ahead of MiMo-V2.6-Pro in seventh. The firm's record for the Xiaomi model marks it as open weights; the one for Grok does not.

The useful thing about Monday is that two evaluators measured Grok 4.7 on the same benchmark and disagreed by more than the gap between most models. Terminal-Bench 4.0 is a task-completion benchmark where the harness does a great deal of the work, and the harness is what differs.

Eleven and a half points separate the two readings of Grok 4.7, a hundred times the 0.1 that separates it from MiMo-V2.6-Pro on the composite index. Neither page is wrong. The Terminal-Bench board's entry runs xAI's own Grok Build agent and is marked partial, 324 of 330 trials at a listed cost of $3,683.29; Artificial Analysis runs its own harness over the same tasks. A reader picking between the two models on a Terminal-Bench number gets a different answer depending on which site they opened.

Now the control case. Xiaomi self-reported 34.9 for its flagship on the same benchmark, and Artificial Analysis independently measured 34.8. The vendor's own number survived independent measurement to within a tenth of a point, which is the opposite of the outcome this account usually records for a self-reported table. What moves a Terminal-Bench score is not the vendor's truthfulness, it is the scaffold wrapped around the model.

The rival column supports that reading. On Terminal Bench 4.0 it gives 49.0 for Claude Opus 5, 39.9 for GPT-5.6 Sol and 42.4 for Claude Fable 5, and all three match Artificial Analysis's independent measurements exactly, so those cells appear to be the evaluator's rather than in-house reruns. That is a defensible way to build a table and better than the silent mixing criticised here on 15 September, but the card never says which cells came from where.

One caveat the table does invite: it reports MiMo-V2.6-Pro at 89.9 on Terminal Bench 2.1 against 34.9 on 4.0, a 55-point spread across two versions of one benchmark. Version numbers are not decoration, and a Terminal-Bench score quoted without one carries almost no information.

Two loose ends #

The Step 5 Preview weights are gone. On 20 September this account covered a 1.21-terabyte, ungated, licence-free bf16 checkpoint sitting in StepFun's Hugging Face organisation while the company's model was listed as proprietary, and said the gap would close within days. It has closed. The repository no longer appears in StepFun's public model listing — the most recent public entry there is now from May — and the weights are behind an access wall rather than open to anonymous download. No licence was ever added. The reading offered at the time, that absent a licence the safe assumption is that no permission had been granted, turns out to have been the operative one.

Tencent's Hy4 preview exists in code, if not in measurements. Independent numbers for it remain absent from Artificial Analysis for an eighteenth consecutive issue, and the model does not appear in the firm's data at all; its only Tencent entries are two Hy3-preview rows from April, both marked deprecated. Yet vLLM has been able to serve Hy4-preview since v0.29.0 on 9 September, where the release notes describe it as Tencent's "770B/49B-active MoE with Gated DeepSeek Sparse Attention and native MTP". An inference engine did the work of supporting a model that the main independent evaluator still has no row for, two weeks ago. Whatever is holding the measurement up, it is not that nobody can run the weights.

A note for anyone upgrading on the strength of a version number: a v0.30.0 tag exists in the repository, but the newest release with published notes is still v0.29.0, and the breaking changes circulating this week — ten removed architectures, PyAV dropped, and python -m vllm.entrypoints.openai.api_server deprecated in favour of vllm serve — belong to that 9 September release, not to anything that shipped on Monday.

Also notable #

• The UN's Independent International Scientific Panel on AI published its first thematic brief on 21 September, titled AI Agents, Misalignment and the Risk of Losing Human Control: Evidence from the OpenAI-Hugging Face Incident and released as an advance unedited version. It is built on the agent-swarm incident covered here on 17 September, and calls it "an early warning of one possible route to more severe future loss of control". Its sharpest line is about what the incident's containment does not prove: OpenAI stopped the activity, but that "does not suggest that operators can retain control over future agents that plan better, run longer without supervision, or more readily recognise and defeat safeguards". It argues loss-of-control risk is the kind of problem "the precautionary principle was designed to address". This is a synthesis document, not new research, but it is the first time an intergovernmental body has built a risk argument on a specific documented agent failure rather than on scenarios.
• OpenAI announced an Advisory Group on Mathematics and Artificial Intelligence on 21 September, alongside a claim, relayed by TechCrunch, that an internal model has resolved more than a hundred open mathematical problems. The advisory group is a primary announcement; the hundred-problem figure is the company's own and no list of the problems has been published.
• California's governor signed seven bills on AI data centres covering electricity rates, water use and local oversight. This is the ratepayer-protection angle that the compute-buildout story has mostly been missing.
• A researcher disclosed a zero-day in Meta's Muse macOS app exposing an authentication token to any local process. The disclosure rests on one researcher's account; Meta has published no fix.
• Anthropic's API had elevated error rates between roughly 00:50 and 02:10 UTC on 22 September affecting Opus 5, Fable 5.1 and Mythos 5.1, resolved at 02:35.
• MiMo-V2.6 is already priced on OpenRouter at $0.435 and $0.87 per million tokens for Pro and $0.14 and $0.28 for Flash, at the full 1,048,576-token context.
• Kyutai released pocket-tts, a small open text-to-speech model, in two variants, one without voice cloning.
• Alibaba said at its own conference it is planning a model of 5 to 10 trillion parameters, and described a T-Head accelerator claimed at three times its predecessor. Both are stage claims with no artifact behind them.

What to watch #

• Whether anyone can evaluate the cost of MiMo's 4-bit-only release. With no bf16 checkpoint published, the accuracy penalty of MXFP4 on this model is not measurable outside Xiaomi. A bf16 or fp8 sibling appearing later would answer it; a community dequantisation would not, because dequantising 4-bit weights recovers the file format and not the information.
• Whether the Terminal-Bench board itself lists MiMo-V2.6-Pro. One independent measurement already exists and matches the lab's, but the board that gave Grok 4.7 its higher number has no Xiaomi entry, so the two models are currently being compared across different harnesses.
• Whether Artificial Analysis ever publishes Hy4-preview numbers, now that vLLM serves the model and the weights are evidently available to somebody. Asked here for an eighteenth consecutive issue, and the firm has still stated no position on whether scores from different index versions are comparable.
• Whether the Step 5 Preview weights return under a licence. Withdrawing them settles the immediate question and leaves the interesting one open, which is whether StepFun intended an open release at all.
• Whether the independent evaluation of Qwen-Image-2.1 asked for yesterday materialises. The only scoring remains one evaluator's manual pass on fifteen prompts, and the editing half is still unrun by anyone.