Digest, issue 1Weights Not WallsDraft
Open weights digest, issue 1
Four new reviewed models in the week to 19 November, two of them from Tencent at about 220B parameters. No new independent results for open models.
This is the weekly digest of Weights Not Walls: what was published in open weights this week, what changed in the scores and prices, and news from the fund and the community. It covers 13 to 19 November 2026. Every fact below comes from our data snapshot of 19 November; the model pages show the sources.
New this week
Four new models from the makers we review were published on the Hugging Face Hub this week. Two are small: both fit an 8 GB graphics card. The other two, from Tencent, are about 220B parameters each, and only one of them fits any of the four machines below, and only in a heavily compressed form.
Each machine column gives the largest format that fits with an 8k context: "yes" fits, "tight" fits with little room to spare, "with RAM offload" runs partly from system RAM and slowly, "no" does not fit. "(estimated)" means no measured file yet. How we work this out: the method.
| Model | Maker | Size | Licence | Published | Graphics card, 8 GB | Graphics card, 24 GB | Apple silicon Mac, 32 GB | Apple silicon Mac, 128 GB |
|---|---|---|---|---|---|---|---|---|
| Aion-1.0-Instruct-2.3B | microsoft | 2.3B | permissive | 2026-11-18 | yes, BF16 | yes, BF16 | yes, BF16 | yes, BF16 |
| SenseNova-Loop-9B | sensenova | 9.4B | permissive | 2026-11-17 | yes, IQ4_XS | yes, BF16 | yes, BF16 | yes, BF16 |
| Hy4-flash-base | tencent | 219B | open with conditions | 2026-11-17 | no | no | no | no |
| Hy4-flash | tencent | 220B | open with conditions | 2026-11-17 | no | no | no | yes, IQ2_XXS |
Aion-1.0-Instruct-2.3B from Microsoft. 2.3B parameters under the MIT License (permissive). Context length 32,768 tokens. Small enough to run at full 16-bit precision on all four machines. There are 4 quantised repos on the Hub so far; we measured the files in 2 GGUF repos. No API provider in the Hugging Face router serves it yet.
SenseNova-Loop-9B from SenseNova. 9.4B parameters under the Apache License 2.0 (permissive). It accepts images as well as text. It is gated: you accept the maker's terms on the Hub before you can download it. An 8 GB graphics card holds the IQ4_XS version; 24 GB holds the full 16-bit weights. There are 42 quantised repos on the Hub so far; we measured the files in 2 GGUF repos. To use images with a GGUF file you also need the vision projector (0.86 GB), which the sizes above leave out. No API provider in the Hugging Face router serves it yet.
Hy4-flash from Tencent. 220B parameters under the Tencent Hy License, Version 1.0 (open with conditions: the model page has our reading of the conditions). Context length 1,048,576 tokens. It accepts images as well as text. Of the four machines, only the 128 GB Mac can hold it, and only as IQ2_XXS (78.9 GB), a format of 3 bits or fewer per weight that gives up more quality than the usual 4-bit versions. For images you also need the vision projector (0.91 GB), which that size leaves out. There are 22 quantised repos on the Hub so far; we measured the files in 2 GGUF repos. By API, Tencent (the maker) serves it at fp8, at $0.035 per million input tokens and $0.14 per million output tokens.
Hy4-flash-base from Tencent, published the same day. Tencent tags it as pretrained, where Hy4-flash is tagged instruct. 219B parameters, the same licence and the same context length, and it also accepts images. There were no quantised versions on the Hub when we fetched the data, so its sizes are estimated from the parameter count. None of the four machines can hold it with an 8k context. No API provider in the Hugging Face router serves it yet.
Scores that changed
No new independent results for open models in the data this week. We count a result when Epoch AI's data gives it a run or evaluation date in the week; collected results without a date are not counted.
Epoch AI's data has 17 new results dated this week, all for closed models. Closed models appear only as reference rows on the benchmark pages. The most recent independent result for an open model in the data is dated 6 November 2026.
Prices
We do not keep price history yet, so there are no price changes to report. This section starts once we keep dated snapshots of provider prices. Current prices per provider are on each model page.
Worth reading
This section is for links chosen by an editor. There is no editor yet, so this example issue has none, rather than links picked by a script.
From the fund and the community
Nothing to report. The fund is not open and the community programmes have not started. When either does, it will be here.
Model data from the Hugging Face Hub. Independent scores from Epoch AI, 'Capabilities & benchmarking' (CC-BY 4.0), with each result's own source linked. Sizes and fits are worked out by Weights Not Walls from the files on the Hub.