At 1 million tokens, Xiaomi reports 2.69 GB of KV cache for HiSparse 2, compared with 12.09 GB for Hybrid SWA. Its test of whether the model can find the right information stops at 256k tokens. Those results answer different questions.
HiSparse 2 Keeps Less of a Long History
An agent might run a short command and receive a long test log. Its next step has to process that growing history, even if only a few lines matter. HiSparse 2 is Xiaomi’s attempt to make that work cheaper while still finding the relevant information.
The design reuses stored information across layers. Upper layers draw on information from lower layers, while sparse layers inspect selected tokens instead of keeping their own full copies. Only 5 of 49 layers use full attention, down from 9 in Xiaomi’s earlier Hybrid SWA design.
On Xiaomi’s 80 billion parameter research model, KV-cache use at 1 million tokens was:
- HiSparse 2: 2.69 GB
- Original HiSparse: 6.72 GB
- Hybrid SWA: 12.09 GB
Xiaomi also calculated 5.02 times less prefill compute than Hybrid SWA and 2.92 times less than original HiSparse. These are counts of operations, not measurements of elapsed time. The cache comparison says how much stored information each design used at that length; the compute figures describe a different part of the work.
Retrieval Was Tested at 256k
Saving cache would be less useful if the model lost track of the information it needed. In Xiaomi’s reported 256k-token RULER/DAGI 2 test, HiSparse 2 scored 58.45. Original HiSparse scored 32.61, and Hybrid SWA scored 35.74.
That result is encouraging within the test Xiaomi reported. HiSparse 2 used less cache at the longer length and scored higher than both comparison designs in the shorter retrieval test. But the retrieval test did not reach the 1 million-token length used for the memory and compute results.
MiMo-v3 weights have not shipped. For now, the reported figures leave one question unanswered: does HiSparse 2 still find the right information when the context reaches the full length of its cache test?
Opus 5.5’s Coding Results Depend on the Checks
The same need to read results at their tested scope appears in the Opus 5.5 coding reports. Anthropic says Opus translated the C load balancer HAProxy to Rust in 9 and 1/2 hours; Fable 5.1 took 12 hours. Both passed nearly all of HAProxy’s regression tests, and Anthropic says Opus cost 51% less.
Those are concrete results for the work and checks Anthropic described. Anthropic did not identify the failed tests or say either rewrite was ready for production. “Nearly all” is useful information, but the missing test details matter when judging a rewrite.
The known-bug tests were mixed. Across 80 bug patterns, CodeRabbit scored Opus 51 against 49 for its baseline. On 13 hard cases, the scores were 10 against five. Yet Opus missed nine bugs the baseline caught and used roughly 40 to 60% more tokens.
GWD’s Hacker News test gives another narrow comparison. Opus found eight of 14 known issues for $15.40 in equivalent API cost, while Fable found seven for $66.34. It was a small test, and Opus still missed six issues. The price and issue count make the result interesting without settling what a different codebase or set of checks would show.
Browser Builds Have Inputs and Fixes, Too
Spearliss turns a photo into a continuous-line drawing in the browser. After an initial model download of about 54 MB, it works offline without uploading the photo. Its repository credits Claude Code but does not name the model, so the build cannot be tied to a specific model from that credit.
Ryan Sale reported making an interactive camera lens lab in one run, taking 1 hour 26 minutes and $25.66 at API rates. The prompt also used files from earlier unpublished projects and maximum effort. Those inputs are part of what “one run” means here.
A repository documents three browser games made from single-sentence prompts, and its pelican game needed a fix for an intermittent black screen. Higgsfield showed a 30-second house-photo-to-3D demo without disclosing the number of attempts or the manual work behind it. The visible builds are real parts of these reports; so are the stated inputs, fixes, and gaps in what was disclosed.
For the long-context work, the next useful result is clear: will HiSparse 2’s retrieval score hold at the full 1 million-token context length?