On Quen 3.5 27B, reported two-bit accuracy rose from 60.45% to 70.96%. Full precision scored 77.10%. The compressed model got substantially better, but the remaining gap is still part of the result.
That distinction runs through the other reports here. Passing a cache between models beat a written hint in a limited test, while a lower-priced coding model prompted discussion of how developers might route work. The gains are specific; so are the limits around them.
Two Bits Improved, but Full Precision Still Leads
At two bits, each weight has just four possible values. That saves space, but answers can suffer. A paper covered by Cloud-Codes changed how Omniquant tunes the compressed model: its measure, RMSSE, keeps blocks with large errors from getting much stronger training pushes than blocks with small errors.
The reported accuracy gain is easy to see:
- Earlier two-bit result: 60.45%
- Improved two-bit result: 70.96%
- Full precision: 77.10%
Perplexity tells a similar story. On Quen 3.5 27B, the reported two-bit figure fell from 37.51 to 22.20, where lower is better. Full precision scored 6.90. RMSSE narrowed the gap, but it did not close it on either measure.
The change did far less at four bits: perplexity moved from 7.17 to 7.09. The paper also reported little added cost while making the compressed file. Runtime ranged from 1.19% lower to 2.82% higher, with 0.00% more peak GPU memory. Those figures describe the file-making step, alongside the reported quality results.
The 8 GB Question Is Still Open
Does that two-bit result mean Quen 3.5 27B will work well on an 8 GB GPU? The reported tests do not answer that. The Cloud-Codes speaker estimates 6.75 GB for the two-bit weights alone, before extra file data, and none of the results came from an 8 GB GPU.
The speaker personally favors a 9B model at four bits for that card. That is a preference, separate from the paper’s Quen 3.5 27B measurements. Code and checkpoints are still listed as coming soon, so the reported two-bit result cannot yet be reproduced from them.
The accuracy gain is substantial within the reported test. Keeping the full-precision scores and the missing reproduction materials beside it gives a clearer picture than the improved two-bit number alone.
Passing a Cache Beat Writing a Hint in Limited Tests
Another paper covered by Cloud-Codes tested C2C, or cache to cache. One model passes another the numbers it saved while reading a question, without writing a hint one word at a time. The comparison asks whether that handoff can carry something useful to the receiving model.
In the paper’s Open Book QA test, the receiver scored 52.6% when the sharer sent no words. When the sharer wrote a sentence instead, the receiver scored 44%. For the same model pair on that test, C2C beat the written route by 8.6 points.
The timing needs its own scope. In one setup, writing the roughly 80-token sentence took 1,312 milliseconds, while cache fusion took about 90 milliseconds. Across settings, the paper reported an average speed difference of 2.5 times. The striking single setup was not the average result.
C2C also needs a trained fuser for each model pair to translate between caches. The tests used multiple-choice questions and models from 0.6 to 14 billion parameters. The paper says an inaccurate sharer can mislead the receiver, and these tests do not show how C2C handles long answers.
A Cheaper Model Still Needs a Useful Workflow
AI Tweets reported that Claude Haiku 5.5 arrived in Claude Code and the Claude Platform at about 75% less cost than Haiku 4.5. In Cursor, the company says short requests with Haiku 5.5 cost about 10× less than with Haiku 4.5. Those are different reported price comparisons, with different scopes.
Developers described a possible workflow: send routine work to a smaller model, then call on a larger one when needed. The price difference makes the proposal concrete, but the script gives no measured saving across projects. A lower request price alone does not establish what a whole project would cost.
NateBJones makes the product side tangible. After compiling 26 agent use cases, he says friction increasingly comes from the product around the model. In his own workflow, he used Dots to combine financial sources and email into a subscription spreadsheet, then gave the spreadsheet to Muse to help act on it. He describes Dots as slower but good at combining information, and Muse as faster at navigating the web and doing tasks.
These are useful results at their stated sizes: a better two-bit score, a faster cache handoff in the reported tests, and lower prices for specified requests. The two-bit code and checkpoints are still pending. C2C’s tests still leave long answers, including cases where the sharer is wrong, open.