10GB of memory, minus 4.9GB of 4-bit weights, leaves 5.1GB.
That remainder has to hold the cache for a 512K context and the inference framework itself.
SenseTime's hardware note for the production release of SenseNova Loop-9B is one line: under 10GB.
One card, one copy
The specifications: about 9 billion parameters, a Loop architecture, context up to 512K, and a weight file of 4.9GB after 4-bit quantisation.
Put together, that is a machine you can carry: one 10GB card, one file for an external drive, no server room.
A 9B model does not win the argument about capability on the parameter table. It wins a different one, about where the model can run.
The parameter table describes a ceiling. Memory describes a floor. The floor decides who gets to use the thing.
512K of context is a second threshold. A contract of several hundred pages, or a mid-sized repository, goes in at once without chunking first. Chunking splits a document and reassembles an answer; long context keeps it on one desk, pressing harder on the cache than on the engineering.
Nothing gets compared until it runs.
Two scores, opposite directions
On two evaluations, SenseTime published the competitor's numbers beside its own. NL2Repo: 41.36 against 44.60 for Qwen3.8-27B. AIDABench: 65.31 against 65.01.
One gap is 3.24 points. The other is 0.30 points, in the other direction.
The claim in the title refers to the second.
How those scores were produced is not fully disclosed: prompts, number of runs, mean or best-of, whether the competitor's figure was run in-house. Without that, a score is a reference point, not a ranking.
So read the two numbers separately.
Three-tenths is not a win
65.31 against 65.01 is a gap of 0.46%.
Call it a tie, and the tie is the finding: three times the parameters bought no measurable advantage. 0.30 points does not show 9B is stronger. It shows the 27B advantage disappeared, and disappearing is not reversing, though the headline claims it is.
Settling it would take more benchmarks. Only two are public.
The loss is real
41.36 against 44.60 is 7.3% below the base of 44.60.
That is not noise. Prompt changes move a score by a point or two; 3.24 is beyond that.
NL2Repo tests repository-level coding: long inputs, many dependencies, edits scattered across files. Trailing there means 9B is still short on long chains of code operations. The direction matches the parameter count, since 9B is one third of 27B.
Publishing the losing number is the more informative choice. A launch that reports only the win is the norm. Reporting both limits "matching" to one benchmark instead of letting it harden into a general claim.
Loop reuses the same layers
SenseTime's description runs to one clause: the Loop structure reuses layer parameters.
How many layers, how many loops, is not disclosed.
What can be confirmed is the outcome, a 9B budget inside a 27B score band. The name describes the mechanism: the same layers applied repeatedly, depth from repetition rather than width.
The saving is in parameters. It is not in compute. Every loop is another pass, so a smaller table can still produce a similar bill.
Loop also differs from distillation and quantisation, which compress a model already trained. Loop places fewer parameters by design and trains them, so the ceiling is set during training rather than inherited.
The subtraction that decides deployment
The first gate in deployment is not a benchmark. It is memory.
A 9B model at 4 bits is 4.9GB. A 27B model at the same precision works out to about 13.5GB, or 2.75 times as much.
That is before the context cache, which grows with a 512K context: the less the weights take, the more room for the document. Whether one machine holds both is a subtraction, not a score.
On that subtraction, 9B starts 8.6GB ahead. The same card that forces a 27B deployment to ration context lets a 9B deployment ask how many instances it can run.
SenseTime's memory figure is "under 10GB", with no statement of the context length at which it was measured. At a full 512K the number rises, by an undisclosed amount.
Revenue beside the benchmark
SenseTime reported 3.79 billion yuan of revenue in the first nine months, up 30.4%. How much of it the 9B model contributes is not broken out.
Shrink the model until one card runs it and you are selling a delivery, not a call. Government and enterprise buyers cannot absorb a data centre. They can absorb a workstation.
Per-call revenue scales with usage, and so does its cost. Per-box is priced when the contract is signed, and every extra machine adds deployment labour.
Phai-A1, released on November 18 by a Hong Kong-listed vendor (6682.HK), also lists local deployment on a personal computer. Its own figure says a coding agent on a single RTX3090 responds 8 to 10 times faster than Qwen27B. That is a vendor claim, without third-party replication. What such claims omit is not the multiple but the conditions: precision, batch size, input length.
Time the benchmark does not measure
Beyond the two scores sits a number nobody benchmarks: the time from a 4.9GB file to a first token. Copy the weights, install a framework, edit a few lines of config.
A 27B model takes more steps, each needing attention. That time never appears in a results table but does appear in a purchase decision, where whether it runs at all comes before whether it runs fast.
Vendors can supply weight size and a memory floor. The rest is a trial.
Four numbers, 4.9GB, 512K, 41.36 and 65.31, put the two models 0.30 points apart on one benchmark and 8.6GB apart on memory. On this road, the parameter table is no longer the first page anyone reads.
*Disclaimer: This article is based on publicly available information and is for reference only. It does not constitute investment advice.*
