Tessera
Status

Where Tessera is, in plain text.

The state of the pipeline, the studio, and the road ahead. Honest about what is built, what is qualifying, and what is on the list.

Shipped

What is in the box today.

Released — in the box

  • l1-l6 spine — The six-layer runtime-aware measurement loop (v3.1 spec): trigger-gated row capture with a true FP16 reference (L1/L1.5), activation-space differential with spectral norm + effective rank (L2), KL divergence with four-forward attribution (L3), domain-weighted prompt bank with per-layer flag telemetry (L4), Hessian sensitivity scorer with Cholesky + dispatch (L5), tail-weighted quant loss with acceptance verdict (L6).
  • per_tensor_calibrate — Offline Frobenius; multimodal extension in qualifying.
  • awq-evolve — Island GA; fitness against the kernel dequant output.
  • llama-quantize --tessera-mode — End-to-end pipeline; opt back to stock with --tessera-mode=off.
  • l5_orchestrator — Iterative requant with family-aware bit allocation.
  • unified_calibrate — Co-calibrate trunk + drafters (DFlash, DSPark, MTP) in one pass.
  • embedding_budget — Size-envelope producer for shared embedding tensors.
  • tessera-ab-harness — Per-tensor t_l² against the kernel dequant output.
  • tessera-dpace — Adaptive per-position D-PACE loss (weights replace the fixed exponential decay) + DFlash block dataset emission from spec_calib.v2.
  • Granite T640 family — granite-4.1-30b, guardian-4.1-8b, granite-4.0-h-small are first-class T640 targets; granite4vision (text / vision / video), granite-docling, granite-speech-plus, and the modern-bert embeddings convert and load through the singular GGUF pipeline.
  • Nemotron 3.5 — Full BF16 conversion (per-expert FFN + MTP blocks); readback-validated, 6,343 tensors, shapes and dtypes checked against the source.
  • Tessera Studio agent loop — Approval tiers, notification budget, time-limited undo, action audit log, chat dock, citation chips with inline hard-stop (ux-fatigue waves 1A–3D).

Qualifying — on the runway

  • multimodal_calibrate — v1 forward-pass is synthetic; C++ mtmd capture wired.
  • clip-capture — True batching, dead-node detection, audio decode wired.
  • tessera-train-dflash — The drafter training driver; the DFlash drafter is feature-conditioned (it consumes trunk hidden states), so the token block dataset is necessary but not sufficient — the offline feature-capture pipeline is the next link.
Qualifying

What is on the runway.

  • Real C++ multimodal capture (clip-capture v2 is the production-readiness line; the synthetic v1 path is being removed as the real path lands).
  • Drafter training drivers — the LK / autoregressive path is native C++; the DFlash / block-parallel path needs the offline feature-capture pipeline before the weighted-CE training graph can fuse.
  • Regime router (per family + shape, dynamic dispatch to the v2 quant functions) — replacing the static thresholds with a learned policy.
  • Linux GTK4 Studio port — embedded Postgres, data contracts, OpenVINO default, OAuth2 surfaces; the port-gap analysis is closed and the build surface is being absorbed.
Not yet

What is on the list.

  • Auto-build calibration corpus from the user's own usage (privacy-preserving; CC BY 4.0 license, scrubber-first; context windows capped at ≤50k tokens per the context-rot finding).
  • Self-improving drafter training end to end (Qwen-class trunk + DFlash + DSpark, regime-routed; the MVP is the multi-dimensional behavioral eval + capability archive).
  • Hardware-backend reach beyond Apple Silicon (Tessera is macOS / iOS first; the pipeline is backend-agnostic, the Linux Studio port is in progress, and the runtime stays Apple-targeted today).