The model, redesigned.
Built from first principles for the processor it would actually run on — not a quantised, pruned copy of an architecture that assumed thousands of parallel cores and a memory system to match.
Sovereign AI. On a CPU.
Everyone else needs a cluster. We spent twelve months making sure you need nothing you don't already own.
- Runs on the CPUs in your racks today
- No GPU estate, no cluster, no queue
- On-premise, private cloud or at the edge
0Billion parameters, in the class we ship
0GPUs required. Not one.
0Bytes leaving your infrastructure
3Ways to deploy the same backbone
The premise
A frontier model is a memory problem wearing a compute costume.
The industry's answer to that was to buy more accelerators. It works, and it produces a system that only ever lives in somebody's data centre — because that is the only place the hardware exists.
We took the other branch. If the intelligence has to live where the data lives, then it has to run on the machines that are already there: ordinary server processors, in ordinary racks, in ordinary buildings.
That constraint changed every decision that followed.
The queue on the left is the bottleneck. Adding processors widens the right-hand side of a road that was never narrow there.
What makes it possible
What twelve months without shortcuts buys you.
Neurons that rest.
A query does not need every parameter in the network. Only the experts relevant to the token do the work; the rest stay quiet, and the arithmetic you skip is the arithmetic you never have to buy hardware for.
Data movement, rewritten.
On a CPU the bottleneck is rarely the maths — it is getting weights to the vector units in the right order. We laid the execution out for cache hierarchies and memory channels instead of for warps.
Portable by construction.
Commodity x86, a private cloud instance, or a browser tab over WebGPU. No CUDA. No exotic silicon. No procurement cycle for parts with a twelve-month lead time.
The how stays ours. Full technical detail is available under NDA.
The evidence layer
Benchmark the deployment. Not the brochure.
A leaderboard measures capability. It does not tell you how the model behaves after quantisation, on the CPU in your rack, with your context length and concurrency.
So the public record has two halves: standard suites that can be compared, and serving measurements taken on the hardware that actually ships.
SC-EVAL · PUBLICATION GATE
4 released results
-
Reasoning
Exact-match accuracyMMLU-Pro · 280-question balanced audit
Twenty deterministic questions from each of 14 domains. This direct-answer audit is derived from MMLU-Pro; it is not a full official leaderboard submission.
- Genie 1 58.21% 163/280 · repeat 58.21%
- Genie 2 64.64% 181/280 · repeat 64.29%
Reproduced · 06 Aug 2026 64.64% Best · Genie 2 -
Repeatability
Repeat exact-match accuracyMMLU-Pro · independent reproduction
The complete 280-question audit was run again against the same pinned builds. Genie 1 repeated exactly; Genie 2 moved by 0.35 percentage points.
- Genie 1 58.21% 163/280 · spread 0.00 pp
- Genie 2 64.29% 180/280 · spread 0.35 pp
Reproduced · 06 Aug 2026 64.29% Best repeat · Genie 2 -
On-box performance
Median decode tokens/sSingle-stream decode · 5 × 256 tokens
One warm-up followed by five fixed, non-streamed prompts at concurrency one. The score is the median server-reported decode rate.
- Genie 1 23.06 tok/s mean 23.17 · range 22.18–24.99
- Genie 2 15.66 tok/s mean 15.62 · range 15.47–15.71
Superchat reproducibility protocol Open run recordMeasured · 06 Aug 2026 23.06 tok/s Best · Genie 1 -
Prompt processing
Median prompt tokens/sSingle-stream ingest · 5 fixed prompts
The same serving run also records how quickly each CPU deployment ingests the fixed prompt before generating the answer.
- Genie 1 119.35 tok/s range 113.08–122.98 · 5 runs
- Genie 2 86.09 tok/s range 82.57–90.23 · 5 runs
Superchat reproducibility protocol Open run recordMeasured · 06 Aug 2026 119.35 tok/s Best · Genie 1
A score reaches this page only after
Build pinnedExact model, weights, quantisation and inference engine.
Conditions loggedHardware, context, batch size, prompt and seed.
Outputs archivedRaw responses, evaluator version and failure record.
Result reproducedA second run agrees, or repeated measurements expose the spread.
Hardware used for this release
Genie 1
AMD EPYC 7502P · 32 cores / 64 threads
128 GB · 4 × 32 GB · 3200 MT/s
Superchat Genie 1 · CPU production profile
Genie 2
AMD EPYC 7401P · 24 cores / 48 threads
256 GB · 8 × 32 GB · 2400 MT/s configured
Superchat Genie 2 · CPU production profile
Next controlled releases
CodingLiveCodeBench · SWE-bench Verified Run pending Tool useBFCL V4 · multi-turn function calling Run pendingRelease SC-EVAL-2026-08-06 names the model files, checksums, engine commits, prompts, seed, raw outputs and both reproduction runs. The MMLU-Pro number above is a transparent 280-question audit, not a full official leaderboard submission.
One backbone. Three deployments.
One product family, wherever your risk register lets it live.
On-premise hardware
A box at your site.
Dedicated hardware, racked in your building. Unplug the ethernet and it keeps answering. The deployment of record for defence, government and critical infrastructure.
Private cloud
Yours. Not theirs.
A single-tenant instance inside your own cloud account and region. It never traverses the public internet and it is never shared with another customer.
On-device browser
In a browser tab.
WebGPU inference on the end user's own device. Once the browser deployment is loaded, prompts and responses remain on that device.
One model, three modalities
Voice, text and vision — all of it local.
Every token, every pixel and every syllable is processed on the customer's own processor. There is no fallback path that quietly sends the hard cases to a cloud.
Everything inside the ring is your hardware. There is no line leaving it.
Speech in, speech out, 12+ languages
Recognition and synthesis both run on device, Indic languages included. Call audio never becomes somebody else's training data.
Reasoning, drafting, translation
The same 120B-class backbone that powers the enterprise assistant and the coding agent.
Documents, charts, scenes
Understanding scanned paperwork, reading charts and interpreting a camera feed, without the image leaving the building.
What your security team will ask
The answers are unusually short.
Nowhere. Inference happens on infrastructure you own, inside your own network boundary.
No. There is no telemetry path that could carry it to us, and on an air-gapped install there is no path at all.
Ours cannot affect you. The model is not reaching us to answer a question.
The deployment is on your hardware, and it keeps running. A licence check calls home once a day to confirm your licence and AMC are current — it carries no prompt, no output and no customer data, and it is never in the path of an answer.
Superchat Research · standing lines
The breakthrough is built. The advantage keeps compounding.
CPU-native inference, conditional computation and speculative decoding are already in production. The work below extends that deployed advantage across performance, efficiency and embodied control.
We publish deployment evidence rather than vanity rankings. Research counts when it reaches a build a customer can run — and then makes the next build faster, smaller or harder to reproduce.
CPU-native inference
In productionWhat does a transformer look like if the target was never a GPU?
Execution laid out for cache hierarchies, memory channels and vector units instead of for warps and shared memory. Most of the work is scheduling and layout, not arithmetic.
Conditional computation
In productionHow few parameters can answer this token well?
Routing each token to the smallest competent set of experts, and keeping the routing itself cheap enough that it does not eat the saving it creates.
Speculative decoding
In productionCan a slow machine feel immediate?
Draft ahead, verify in bulk, discard what the full model rejects. Correctness is unchanged by construction; only the wall-clock changes.
Embodied control
Active · SwayamCan the same backbone close a balance loop in milliseconds?
Language and planning at the top, joint-level correction at the bottom, one model between them — running on the robot, with no link to lose.
Methodology and measured results are shared under NDA. We do not publish numbers we have not had checked.