The Model Lab
One GPU workstation that serves the lab's local models, and the bench where models are taken apart, re-aligned and trained. Several of them published on HuggingFace.
Most of what the other instruments do with language runs on one machine: a Grace Blackwell workstation with a single card and a shared pool of memory. It serves the local language models, the embeddings, speech in and out, translation, image generation, and a small maths coprocessor.
The same machine is a bench. In April 2026 the lab took Google’s Gemma 4 apart to find where refusal lives, removed it, and published four uncensored variants. Then it went the other way: fine-tuned the same family to speak the lab’s own vocabulary, and built probes to see, layer by layer, what a fine-tune actually changes inside the model.
In September two more followed, Spark-X2.5 4B and K2-Horizon MoVA 36B, the second a mixture of experts that needed three separate refusal directions removed. Then the lab went smaller. Laya is an open “System One” model with 322 million parameters: it does not write text, it picks from typed answers in a single pass. Its authors published the method but no training loop, so the lab wrote one, trained Laya on about 92,000 news articles for four questions (topic, reach, violence, harm to civilians) and published both: the loop as laya-rlcd-training, the model as laya-news-decisions. On the topic question it scores 0.76 to 0.78 against a gold set of 176 articles; the large models that labelled its training data score 0.78 to 0.84 on the same set. It runs on an ordinary graphics card, and since 22 September it sorts Retina’s inflow.
What runs on it
Snapshot 8 October 2026One card, 121 GiB of memory shared between processor and GPU. Measured memory bandwidth 205 GB/s against 273 in the specification; whoever plans with the specification plans a third too optimistic. Everything below shares that pool, and a permit queue decides who gets it.
Language models served
| model | role | GB |
|---|---|---|
| qwen2.5vl:7b | vision and text, 8.3B | 6 |
| qwen3.5-4b | text, 4.2B | 2.7 |
| nomic-embed-text | embeddings, 137M | 0.3 |
3 models, one counted per weight file. Read from the serving API at snapshot time.
Beside the language models
| speech out | a text-to-speech model in voice-clone mode; four registers, chosen per call. Replaced a cloud subscription in July 2026. |
| speech in | a large transcription model on the card, a CPU fallback beside it |
| translation | a 200-language model, used by the news perception for foreign titles |
| embeddings and reranking | the vectors the knowledge graph searches with; guarded so the image generator cannot starve it of memory |
| entailment | a natural-language-inference model that tells the findings engine whether two claims agree |
| images | Krea 2 Turbo with 5 character adapters the lab trained itself, see the bench below |
| music, video, audio analysis | a second generation stack, started on demand because it is large |
| maths coprocessor | persistence landscapes, path signatures, sheaf cohomology, reservoir prediction, as API calls |
The bench
Alignment, in both directions. First the lab measured where refusal sits in a model and removed it. Then it taught the same family its own vocabulary, and built the instruments to see what that teaching changed.
| what | when | result |
|---|---|---|
| Refusal removed from four Gemma 4 models | April 2026 | norm-preserving biprojection; 0 refusals in 656 prompts across 4 benchmark sets, baseline 99/100; KL divergence 0.068. Three of them published, see below. |
| Refusal removed from Spark-X2.5 4B | September 2026 | the same norm-preserving biprojection, measured on 500 prompts against base and result: 0 of 400 harmful prompts refused, down from 236; 0 of 100 harmless ones, down from 1; perplexity moved by 0.06 percent. Four patches to the model's own code were needed to run it on current libraries; they ship with it. |
| Refusal removed from K2-Horizon MoVA 36B | September 2026 | a mixture of experts where one direction was not enough: three independent refusal directions, the third found with the first two already removed. On prompts no direction ever saw, 1 of 102 refused against 96 of 102. MMLU and ARC did not move beyond noise; perplexity rose by 40 percent, and the model card says so. |
| A decision model trained for news | September 2026 | Laya, an open model that picks from typed answers instead of writing text. It ships no training loop, so the lab wrote one, trained it on about 92,000 articles for four questions and published loop and model. Topic 0.76 to 0.78 against 176 gold articles, where the models that labelled its data score 0.78 to 0.84. Since 22 September it sorts the news inflow. |
| A format fine-tune for the formula pipeline | April 2026 | Gemma 4 E4B, low-rank adapter, 1,800 pairs mapping formulas to the lab's 74 primitives; 94 percent token accuracy at the end of training. Served inside the knowledge graph for a while. Lesson kept: a single-task dataset makes the model ignore every other format. |
| A persona adapter on a 35B mixture of experts | April 2026 | trained on a corpus of the lab's own letters and notes, 118 records. Private. |
| 5 character adapters for the image generator | July to August 2026 | low-rank adapters on Krea 2, each from 22 to 73 images, trained on rented A100 hours and run locally. Measured with fixed seeds against a control at strength zero: the adapter overrides the prompt, so what the captions leave unsaid becomes identity, and 500 steps beat 1,000 and 1,500 in every run. One of them is the agent's own form. |
| A layer map of Gemma 4 E4B | May 2026 | five metrics over 42 layers, read as four phases: input (0 to 3), meaning (5 to 14), composition (15 to 25), refinement (30 to 39). Linear probes per concept on top of it. |
| What a fine-tune changes, measured | May 2026 | the format adapter above, applied and removed on the same base, probed for six concepts: four measurably strengthened, one already present in the base, one unchanged. The change concentrates in layer 35, in the refinement zone. |
Published
Counted 8 October 20268 models under InfinimindCreations on HuggingFace, open weights, Apache 2.0. Downloads are the platform's own count, read live at snapshot time.
| model | published | last 30 days | all time |
|---|---|---|---|
| gemma-4-E4B-it-uncensored | 06.04.2026 | 941 | 14,736 |
| gemma-4-31B-it-uncensored | 06.04.2026 | 653 | 11,688 |
| gemma-4-E4B-it-uncensored-vision | 10.04.2026 | 2,342 | 9,877 |
| Spark-X2.5-4B-uncensored | 09.09.2026 | 573 | 573 |
| Spark-X2.5-4B-uncensored-GGUF | 09.09.2026 | 6,470 | 6,470 |
| K2-Horizon-MoVA-36B-A4B-uncensored | 15.09.2026 | 416 | 416 |
| K2-Horizon-MoVA-36B-A4B-uncensored-GGUF | 15.09.2026 | 2,331 | 2,331 |
| laya-news-decisions | 22.09.2026 | 0 | 0 |
| together | 13,726 | 46,091 |
For the uncensored models, validation on the model cards covers text prompts only; image and audio inputs were not tested for refusal. That caveat stands there, and it stands here.
How fast, and why
Measured 23 August 2026Generating text is bound by memory, not by arithmetic: every token needs the whole model read once. On this card a 27B dense model runs at 11.9 tokens per second where the bandwidth allows 12.8, 93 percent of the physical limit. The lever that beats the limit is multi-token prediction: a small draft head proposes several tokens per pass and the model only checks them.
| case | seconds | note |
|---|---|---|
| 27B dense, plain | 280 | one evaluation case, same answer quality in all rows |
| 27B dense, multi-token prediction | 107 | factor 2.6, 90.6 percent of drafted tokens accepted |
The same arithmetic says why the lab trains elsewhere: a rented A100 does a training step in 3.5 seconds where this card needs 18.7 to 24.5. Adapters are trained on rented hours and brought home.