/inference/gemma4-e2b

gemma4:e2b

Measured by 8 nodes on hardware we do not own, and signed by each of them. What machines here actually proved — never what the card claims.

matrix built 23 Oct 2026 · every field verified against the signature of the node that produced it, and re-verifiable by you: each signed figure carries the exact bytes the node signed (signing_payload_b64) and the signature over them, and the ed25519 public key is embedded in node_did. effective_ctx is the context length a node PROVED by recall, never the length a model card advertises

context declared1,31,072what the runtime reported. A node attests it was told this; nobody attests it is true.

context proved by recall1,31,072the largest size at which a machine could still find tokens planted across the whole prompt — start, middle and end, all of which had to come back.Also served single-stream at 32,768 — the size a machine loaded and answered at, which is a weaker claim than recall and is shown beside it rather than merged into it.

A machine here made it use everything the runtime declared.

recalled 128k▲ served 32k⌐ declared 128k · unsigned

every machine that timed it

4.1 → 170.5 tok/s across 8 machines · ×42.0 spreadunnamed machineNVIDIAx86 · CPUApple silicon

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.

How it was served — proved batch profiles

the same model advertises itself in different shapes, each one a machine proved

Each circle is one signed batch profile; the number above it is how many replies were decoding at once, and the circle grows with the context it was proved at. Up and to the right is better and mostly unreachable — the curve bending down as you go right is the trade: more people served, each of them waiting longer.

4 users, batchedlong-context

32kcontext

27tok/s total

7.9tok/s · per user

4 replies decoding at once — the capacity this gram can actually serve a room.

10.2 GiB resident687 MiB KV per user

did:epn:002408…9e3c79 · q4_0 KVsigned

4 users, batchedhigh-throughput

4kcontext

27tok/s total

7.8tok/s · per user

4 replies decoding at once — the capacity this gram can actually serve a room.

7.8 GiB resident85 MiB KV per user

did:epn:002408…9e3c79 · q4_0 KVsigned

8 users, batchedhigh-throughput

4kcontext

25tok/s total

3.6tok/s · per user

8 replies decoding at once — the capacity this gram can actually serve a room.

did:epn:002408…9e3c79signed

6 users, batchedhigh-throughput

4kcontext

24tok/s total

4.7tok/s · per user

6 replies decoding at once — the capacity this gram can actually serve a room.

did:epn:002408…9e3c79signed

2 users, batchedgeneral

4kcontext

13tok/s total

7.2tok/s · per user

2 replies decoding at once — the capacity this gram can actually serve a room.

did:epn:002408…9e3c79signed

Single streamlong-context

32kcontext

9tok/s total

10.6tok/s · per user

One reply, streaming — the latency a single user sees.

8.2 GiB resident687 MiB KV per user

did:epn:002408…9e3c79 · q4_0 KVsigned

4 users, batchedhigh-throughput

4kcontext

5tok/s total

1.6tok/s · per user

4 replies decoding at once — the capacity this gram can actually serve a room.

7.8 GiB resident85 MiB KV per user

did:epn:002408…2e5c02 · q4_0 KVsigned

Single streaminteractive

4kcontext

3tok/s total

3.3tok/s · per user

One reply, streaming — the latency a single user sees.

7.6 GiB resident85 MiB KV per user

did:epn:002408…2e5c02 · q4_0 KVsigned

Single streaminteractive

4kcontext

2tok/s total

10.7tok/s · per user

One reply, streaming — the latency a single user sees.

7.6 GiB resident85 MiB KV per user

did:epn:002408…9e3c79 · q4_0 KVsigned

Languages, proved

Bengali English Gujarati Hindi Kannada Malayalam Marathi Punjabi Tamil Telugu

Each question was put entirely in its own language and script, and the answer had to be right. A model with no coverage cannot parse the question, so answering in English is a failure here, not a pass. 10 proved of 10 asked — a language absent from the probe set was never asked, and is not a claim either way.

Capabilities, measured

Audioprobed, and failed

Given audio and asked to transcribe it.

Reasons about codepassed

Asked to predict exactly what a short program prints. This is the half of coding that matters for agent work and that writing valid syntax does not demonstrate: following state through control flow.

Writes codepassed

Asked for a specific function and checked with the Go compiler’s own parser. Nothing is executed — running model-written code to score a model would be a security hole opened for a metric — but syntactic validity is a fact rather than an opinion.

Follows instructionspassed

Given a plain constraint on its output format and checked for exact obedience. This separates an instruct-tuned model from a base one — a base model answers fluently and ignores every instruction, which then looks like a dozen unrelated weaknesses instead of one.

Sustained outputpassed

Asked to produce a long structured answer and checked for length. A model that stops after a couple of hundred tokens cannot write a report, however capable it is otherwise.

Structured outputpassed

The ENGINE was asked to constrain decoding to a schema, and the answer conformed. This is stronger than politely asking for JSON and hoping: the grammar makes invalid output unreachable, which is what an operator actually builds on.

Extended thinkingpassed

The model emitted reasoning on the engine’s dedicated reasoning channel before answering. Read from that channel rather than by scanning the answer for tags, which is why models that reason are no longer reported as models that do not.

Synthesises sourcespassed

Given three documents with one fact split across them, and required to combine rather than quote. No single document contains the answer, so a model that retrieves without reasoning cannot pass.

Tool callingpassed

The node asked the model to call a function and checked that it called the RIGHT one with a well-formed argument — not merely that it emitted a tool name.

Faithful to tool resultspassed

The model was handed a list through a tool and had to report every item back. Calling a tool and understanding its answer are different abilities: a model can produce a perfect call, receive nine items, and confidently report three. This is the flag that separates them.

Tool loop finishedpassed

Whether the conversation ENDED after the tool call, or the model kept calling until the turn budget ran out. A model that passes tool calling and fails this will call a tool in production and never come back.

Chained tool callspassed

The model had to call one tool, read its answer, and call a second with a value derived from it. One call is not agentic work — real specialist jobs are sequences, and a model that never takes the second step looks, from outside, exactly like a specialist silently doing nothing.

Counts in imagespassed

Shown a row of shapes and asked how many. Reading rendered text and grasping a scene are different abilities, and a model can do the first with no spatial understanding at all — which is what charts, shelf photographs and document layout actually need.

Multiple images at oncepassed

Given two images and asked to compare them. Many engines and models accept one image per turn and silently drop the rest, which turns a comparison into a confident answer about half the evidence.

Visionpassed

A four-digit code was rendered into an image and the model had to read it back. A model that cannot see cannot guess four digits, and a model that can has no reason to fail — so this passes and fails for the right reason, which describing a picture never did.

Spatial reasoningpassed

Asked which of two shapes is larger and on which side — both ways round, so a coin flip cannot pass. Reading glyphs, counting objects and understanding layout are three different abilities, and only the third answers "what is next to what", which is what charts, diagrams and shelf photographs turn on.

Embeddingsprobed, and failed

The engine was asked for an embedding of a short text and returned a real, non-zero vector.

The gemma4 family

siblings measured on the same network — compare what each actually proved

hf.co/unsloth/gemma-4-E2B-it-GGUF:Q4_K_M/inference/hf.co-unsloth-gemma-4-e2b-it-gguf-q4_k_m

1,31,072 ctx proved8 nodes

hf.co/unsloth/gemma-4-E2B-it-GGUF:Q4_0/inference/hf.co-unsloth-gemma-4-e2b-it-gguf-q4_0

1,31,072 ctx proved3 nodes

hf.co/unsloth/gemma-4-E2B-it-GGUF:IQ4_XS/inference/hf.co-unsloth-gemma-4-e2b-it-gguf-iq4_xs

1,31,072 ctx proved2 nodes

hf.co/unsloth/gemma-4-12b-it-GGUF:Q4_K_M/inference/hf.co-unsloth-gemma-4-12b-it-gguf-q4_k_m

1,31,072 ctx proved1 node

hf.co/unsloth/gemma-4-31B-it-GGUF:Q4_K_M/inference/hf.co-unsloth-gemma-4-31b-it-gguf-q4_k_m

1,31,072 ctx proved1 node

hf.co/unsloth/gemma-4-E2B-it-GGUF:BF16/inference/hf.co-unsloth-gemma-4-e2b-it-gguf-bf16

1,31,072 ctx proved1 node

hf.co/unsloth/gemma-4-E2B-it-GGUF:Q8_0/inference/hf.co-unsloth-gemma-4-e2b-it-gguf-q8_0

1,31,072 ctx proved1 node

hf.co/unsloth/gemma-4-E4B-it-GGUF:Q4_K_M/inference/hf.co-unsloth-gemma-4-e4b-it-gguf-q4_k_m

1,31,072 ctx proved1 node

hf.co/unsloth/gemma-4-E2B-it-GGUF:Q3_K_M/inference/hf.co-unsloth-gemma-4-e2b-it-gguf-q3_k_m

8,192 ctx proved3 nodes

hf.co/unsloth/gemma-4-E2B-it-GGUF:IQ4_NL/inference/hf.co-unsloth-gemma-4-e2b-it-gguf-iq4_nl

8,192 ctx proved2 nodes

hf.co/unsloth/gemma-4-E2B-it-GGUF:Q3_K_S/inference/hf.co-unsloth-gemma-4-e2b-it-gguf-q3_k_s

4,096 ctx proved1 node

signed by this network

What machines here proved

context proved by recall1,31,072

peak throughput ever signed170.45 tok/s · did:epn:002408…84bf54

resident while loaded1.2 GiB (1.2 on card) · did:epn:002408…a5be46

machines that measured it8

samples behind those numbers8

regions it was proved in8

Every line traces to a node signature.

from the public record · unsigned

What the runtime says about it

context declared1,31,072

parameters5.1B

parameter count5,12,31,79,235

quantisationQ4_K_M

familygemma4

formatgguf

ollama /api/show; the node attests that the runtime reported these, not that they are true. Attested by did:epn:002408…9e3c79.The only unsigned block on this page. We attest we read these from the runtime — not that they are true. Nobody signs a model card.

Proven across regions

Bhubaneswar, Odisha752115 · 1 node

The Dalles, Oregon97058 · 1 node

Muzaffarpur, Bihar842001 · 1 node

Hyderabad, Telangana500062 · 1 node

Hyderabad, Telangana500001 · 1 node

Vikarabad, Telangana501202 · 1 node

Bengaluru, Karnataka562130 · 1 node

Sundargarh, Odisha770033 · 1 node

Every machine that measured it

nothing is averaged — the machine is what you are choosing

did:epn:002408…84bf54Bhubaneswar · 752115 · ollama · 15 Sept 2026

0ctx provedno recall at 1,31,072 · 65,536 · 32,768 · 16,384 · 8,192 · 4,096

170.45tok/s now7 samples · 7s

✓ Reasons about code✓ Writes code✓ Follows instructions✓ Sustained output✓ Structured output✓ Extended thinking✓ Synthesises sources✓ Tool calling✓ Faithful to tool results✓ Tool loop finished✓ Chained tool calls✓ Counts in images✓ Multiple images at once✓ Vision✓ Spatial reasoning✗ Audio✓ 10 languages

did:epn:002408…a5be46The Dalles · 97058 · 0.40.6 · on host · 20 Oct 2026Intel Xeon @ 2.00GHz · Tesla T4 15 GiB · 2 cores · 31.3 GiB RAM

1,31,072ctx provedrecalled at every size asked

105.23tok/s nowpeak 121.63 · 20 Oct 202640 samples · 97s1.2 GiB resident · 1.2 GiB on card · 4.3 GiB on disk

✓ Reasons about code✓ Writes code✗ Embeddings✓ Follows instructions✓ Sustained output✓ Structured output✓ Extended thinking✓ Synthesises sources✓ Tool calling✓ Faithful to tool results✓ Tool loop finished✓ Chained tool calls✓ Counts in images✓ Multiple images at once✓ Vision✓ Spatial reasoning✗ Audio✓ 10 languages

did:epn:002408…afbb8aMuzaffarpur · 842001 · 0.32.14 · 20 Aug 2026

8,192ctx provedprobe hZTeG7W7VV…

43.34tok/s now133 samples · 483s

✓ Tool calling✓ Tool loop finished✓ Structured output✗ Extended thinking✓ Vision✗ Audio

did:epn:002408…0c5265Hyderabad · 500062 · 8 Oct 2026

—ctx proved

19.32tok/s now31 samples · 331s

no capability probed

did:epn:002408…0548a8Hyderabad · 500001 · 8 Oct 2026

—ctx proved

19.32tok/s now31 samples · 331s

no capability probed

did:epn:002408…2e5c02Vikarabad · 501202 · ollama · 28 Jul 2026

4,096ctx provedprobe YRyg4OySB6…

8.89tok/s now22 samples · 563s

✓ Tool calling✓ Tool loop finished✓ Structured output✗ Extended thinking✗ Vision✗ Audio

did:epn:002408…9e3c79Bengaluru · 562130 · 0.31.2 · 21 Aug 2026Intel64 Family 6 Model 151 Stepping 5, GenuineIntel · 15.7 GiB RAM

4,096ctx provedprobe PP+vvZ0H4S…

7.72tok/s now778 samples · 23,044s

did:epn:002408…d68672Sundargarh · 770033 · ollama-host · 29 Jul 2026Intel Core i5-5350U @ 1.80GHz · 2 cores · 8.0 GiB RAM

4,096ctx provedprobe m3CV/hFfBj…

4.06tok/s now20 samples · 474s

✓ Tool calling✓ Tool loop finished✗ Structured output✗ Extended thinking✗ Vision✗ Audio

Context ladders attempted: did:epn:002408…84bf54 131072,65536,32768,16384,8192,4096 · did:epn:002408…a5be46 4096,8192,16384,32768,65536,131072 · did:epn:002408…afbb8a 131072,65536,32768,16384,8192,4096 · did:epn:002408…2e5c02 131072,65536,32768,16384,8192,4096 · did:epn:002408…9e3c79 4096 · did:epn:002408…d68672 131072,65536,32768,16384,8192,4096

Run this where your data already is

Tell us the workload, whose machines it should run on, and which jurisdiction it answers to.

Run a gramPut a machine on the network in this region; it signs its capacity into the room.Open a gramxStart your own room — an org, a community, a brand.Be invitedAccept a link from an admin to cross into the private room.