Model labs
Labs inspect a model without changing its weights. Layer importance measures what the blocks do, quantization compares and writes smaller files, the visualizer exposes activations and can design a reusable architecture template, and the latent view shows the hidden geometry.
Layer importance
The lab runs one baseline pass, then one ablated pass per transformer block. The ablated loss is the next-token cross-entropy with that block skipped, and Delta is the increase in nats per token.
- Teal bars are a positive delta: ablating the layer hurt, so the layer matters.
- Yellow bars are a negative delta: ablating it helped on this calibration text.
- The bars are scaled to the worst layer in the run; the absolute delta is in the Delta column.
- The highlighted row is the least important layer, the natural first surgery candidate.
- Attention RMS and FFN RMS show the branch activation magnitude. Very large early FFN spikes are outliers, not importance by themselves.
Calibration text matters. A short, representative prompt is usually more useful than one long sample because every extra token is paid once for the baseline and once per layer.
Surgery and pruned export
The surgery preview skips the least important layers in rank order and shows the loss after each prefix. Rows within 0.05 nats of the baseline are highlighted as low risk.
Each row can write a real GGUF with those blocks removed. The file is renumbered, per-layer metadata is filtered and block_count is updated, so it loads as an ordinary smaller model. The source model is untouched.
Manual surgery lets you name any layers to remove, for example 12, 8, 15. The first and last blocks are kept: they carry the input- and output-adjacent streams. The valid range is 1 .. N-2, where N is the number of blocks.
Quantization
Quantization compares GGUF builds only, so generated .smm models do not appear in its pickers; use the Model visualizer or Latent space for those. The tab has two comparisons. The tensor comparison dequantizes both files and reports relative L2, cosine and worst tensors without a forward pass. It also checks that both files are the same model: same architecture, tensor names and shapes. A pruned or fine-tuned file is not a quantization pair, and the calibration-loss comparison is disabled for it.
The calibration-loss section loads each model in turn and measures the same next-token cross-entropy, so the two models never have to be resident together.
Actual quantization writes a smaller GGUF: a fixed target (Q2_K-Q8_K, Q8_0, Q4_0, F16), a per-role blend (embedding and LM head stay finer), or a byte budget that picks the coarsest K-quant that fits. K-quant rows are written row-padded and the engine pads the input rows at inference, so a model whose hidden size is not 256-aligned (Qwen2-0.5B at 896) still runs the full K-quant.
Model visualizer
The model picker lists both .smm and .gguf files from the Models folder, so a model exported from Training appears after Add to Models.
One prompt runs through the activation hook. Rows are layers, columns are prompt tokens, and the colour is the branch RMS on a log scale. Hover a cell for the layer, token and exact value.
The right-hand column is the layer type from the model config: full or sliding attention, linear for blocks without attention (for example LFM2 short-conv layers), and recurrent for DeltaNet blocks.
The Architecture template section works without a capture. Set the hidden width, depth, heads, KV heads, FFN width and context, then choose full or sliding for each layer. Add layer, Insert and Remove change the depth; the layer list is the ordered plan. Use selected model seeds the shape from the model in the picker, including sliding patterns and non-attention layers. Open template and Save template read and write the plan as JSON. Use in New model sends the dimensions and layer plan to Training, where the same plan can be viewed or edited as a comma-separated list, or loaded again with the Open template button beside Tiny, Small and Medium.
A seeded Gemma-3 model shows its full/sliding pattern even when the GGUF stores a sliding-window period instead of an explicit per-layer mask. Linear and recurrent layers can be kept in the visual plan; the current training kernel trains those as full attention and the editor says so.
Loading and capture are interruptible. The status row shows whether Studio is loading weights or running the forward pass, and Interrupt cancels both stages. Latent space uses the same loading and interrupt controls.
Latent space
The final normalized hidden state of every prompt token is projected to two dimensions with classical MDS, the dual of PCA. Nearby points had similar model states. The scatter labels each point with its decoded token; the legend shows the prompt and the explained variance.
Aggregate keeps earlier prompts in the plot instead of replacing it, with one colour per prompt. Clear plot starts over. Import prompts reads a text file (one prompt per line) or a JSON/JSONL file with prompt, text, content or message fields, then projects the batch with one model load. Export points writes the run, prompt, token, projected X/Y and explained variance as JSONL.
Agents
There are two ways the model can use a specialist. delegate selects one of the saved agents below by name and uses that agent's stored rules, model, tools and permissions. spawn_agent creates a temporary specialist from parameters in the call (name, rules, optional model and tools); it is not saved and starts with Ask permissions. Saved agents are shown to the parent model as an enum of names with their purposes, and the prompt also explains when to delegate and what to pass.
A delegated agent is told it is running one self-contained task for another agent: it cannot see the parent's conversation or ask the user questions, and its reply is a report the parent turns into the final answer.
- Inherit current model agents run on the same weights. Studio runs the agent in its own session so the agent's own tools and permissions still apply.
- Named model agents load that model through the model host for one self-contained task. The host caches it and evicts idle entries after the agent's TTL, or when another load needs the space.
- Rules are the agent's system prompt. Purpose is the one-line description the parent sees.
- Allowed tools can inherit the enabled project tools or be narrowed to a checked set. The toolbox toggles which tools the project offers.
- File and web access is Ask, Allow or Deny per category. A new agent starts at Ask for read, write, delete and web, and an agent's answer never changes the project policy.
- spawn_agent lets the model create a temporary specialist from task-supplied name, rules, model and tools. The temporary agent uses the same Ask defaults.
- Max delegation depth bounds nested agent calls; a delegated agent can call the other enabled specialists until that depth is reached.
Audio and vision
An LFM3-Audio model treats a text prompt as a text-to-speech request: the words are synthesized into the assistant's 24 kHz WAV artifact automatically. There is no Speak button and no separate text reply. The audio card shows its duration and the WAV is saved with the chat.
An attached audio file is transcribed. On an audio model the transcript becomes the assistant reply, and the audio stays on the user message for playback. The backbone alone is not enough: the separate audio projector (mmproj-*.gguf) is required for transcription and TTS. See Audio and mmproj if Send stays disabled after attaching a WAV.
An LFM3-VL model describes an attached image natively. The image card stays on the user message; the caption is the assistant reply. The session options set the image's longest side and the answer token cap.
LFM3-VL is an image-understanding model, not a text-to-image generator. A text-only prompt runs the underlying text backbone; it does not create a picture. To describe or answer a question about an image, attach the image and include the question in the same turn.
Punctuation-only model turns (the audio backbone's bare ":") are treated as no speech and hidden from the transcript.