Measured by 3 nodes on hardware we do not own, and signed by each of them.
Context declared
131,072
what the runtime reported. A node attests it was told this; nobody attests that it is true.
Context proved by recall
8,192
the largest size at which a machine could still find a token planted at the far end.
This model advertises 16× the context any machine on this network could make it use. That is not an accusation — a longer window may work on other hardware, at other quantisation. It is what happened here, and it is signed.
This model calls tools correctly and does not stop. A node watched it call the probe tool, then call it again, until the turn budget ran out. Chosen on “supports tools” alone, it will call a tool in production and never come back.
Quantisation
Q4_K_M
Parameters
5.1B
Family
gemma4
Format
gguf
ollama /api/show; the node attests that the runtime reported these, not that they are true. Attested by did:epn:002408…afbb8a.
Tool calling — passed
The node asked the model to call a function and checked whether it produced a valid call. Passing here is not the same as reasoning correctly over what the tools returned — a smaller model can do the first and fail the second silently.
Tool loop finished — probed, and failed
Whether the conversation ENDED after the tool call, or the model kept calling until the turn budget ran out. Passing “tool calling” and failing this means the model will call a tool in production and never come back. A model chosen on the first flag alone is the reason an agent silently stops delivering work.
Vision — passed
Shown an image and asked what was in it. The node checked the answer.
Audio — probed, and failed
Given audio and asked to transcribe it.
Extended thinking — probed, and failed
The model emits reasoning before its answer. Measured where the probe could observe it.
Structured output — passed
Asked for JSON matching a schema, and got it. This is the difference between a model you can build on and one you must parse defensively.
Nothing below is averaged. A model is fast on one machine and slow on another, and the machine is what you are choosing.
did:epn:002408011220ad2ba1fabbc8e5f23892c0134225790dd035d652ce541dc16e0b098c56afbb8a
Context proved
4,096
Throughput
45.05 tok/s
over 104 samples
Runtime
0.31.2
probe v3
Context sizes attempted: 131072,65536,32768,16384,8192,4096 · measured Jul 13, 2026
did:epn:0024080112203fe69ce44d7b1418dd426331a539f12c1de428a235017442d41ceab0a29e3c79
Context proved
4,096
Throughput
7.65 tok/s
over 446 samples
Runtime
ollama
probe v4
Context sizes attempted: 131072,65536,32768,16384,8192,4096 · measured Jul 17, 2026
did:epn:002408011220841464a4a6d45a8bd729e032b4c1b03d9ad69a58c76e0a5c51bd024bfdd68672
Context proved
4,096
Throughput
4.24 tok/s
over 11 samples
Runtime
ollama-host
probe v4
Context sizes attempted: 131072,65536,32768,16384,8192,4096 · measured Jul 17, 2026