## Model Capabilities

### Context Window

**400,000 tokens** is the fleet-wide per-slot ceiling. The harness clamps max_tokens to `min(45056, contextWindow − est − 4096)` to preserve MTP head and estimation buffer. With 400K context, a 45K token request leaves 355K for MTP head + estimation.

### Multimodal

All three LLM endpoints accept `input: [text, image]`. Vision is native: Radiance and Halogen v2 load their vision towers beside the checkpoint; llama.cpp passes the BF16 mmproj via `--mmproj`.

### Audio

No audio input in the fleet. The FLM NPU endpoint had audio 2026-09-17 → 2026-10-01, but neither MiMo nor Qwen3.5 tower carries it.

### MTP (Multi-Token Prediction)

- **frank-primary** (Radiance): MTP spec decode depth 2, `--max-num-seqs 3`.
- **frank-large** (Halogen): MTP spec decode ALWAYS-ON (`HALOGEN_SPEC_ADAPT=0`), 2 slots × 262,144 ctx, harness window 217,088 = ceiling − 45k estimate headroom.
- **frank-secondary** (llama.cpp): MTP OFF (no `nextn` tensors exist in Qwen3.5-9B Q6_K GGUF).

## Benchmarking Policy

**STOP the production service BEFORE benchmarking**. A live service starves the GPU and corrupts results.

**Procedure**:
1. Stop and mask (self-healing would restart it):
   - Dankclaw `frank-large`: `sudo systemctl stop frank-large && sudo systemctl mask frank-large`
   - Dankclaw `frank-secondary`: `sudo systemctl stop frank-secondary && sudo systemctl mask frank-secondary`
   - Dankserver `frank-primary`: `sudo systemctl stop frank-primary && sudo systemctl mask frank-primary`
2. Bench.
3. Unmask and start; verify `curl http://192.168.0.<host>:7770/health`.

**CRITICAL**: `systemctl disable` deletes the unit file from `/etc/systemd/system/`; `mask` symlinks to `/dev/null` and is safe/reversible. Re-deploy after benching instead of unmasking.

**Gufo bench**: Cannot rank speculative-decode settings; will invert the ranking. Use for shape, not for A/B. Generate **at least 128 tokens** — at 8 tokens, `inter_token` is first-token noise and run-to-run spread swamps real differences. For A/B, replay **real** text through the live server and read the `Server-Timing` header (`ttft;dur=402.485` in the name half, split on `=` keys the dict `ttft;dur`).

**GPU experiment law**: A bare 2-rank RCCL/torch.distributed probe against a live service can wedge the GPU driver (host went hard down ~20 minutes, power cycle required). If an experiment needs GPUs, stop and mask first, or use a host that serves nobody. Distinguish: `ip neigh show <ip>` from a peer — FAILED means off/NIC dead, merely wedged kernel usually still answers ARP.

**Live measurement**: llama.cpp servers report honest per-request numbers in the `/completion` `timings` object (prefill/decode rates, `draft_n`, `draft_n_accepted`).

**Scratch ports**: Not used on LLM hosts — a test process cannot share the single GPU anyway; validate against the service itself.