Free models, counted honestly.

The Court’s recruitment dossier index, built from the Oct 26 research report — now 82 cards, one per distinct weight: grouped family cards were split so every weight stands on its own, deduped across providers (the same weight on five providers is one expert, not five, and its other routes are named on its card). Every route keeps the Free Stuff card format — name, real cost, claim / setup steps, source, and a verified-vs-unverified flag — and every card opens a full dossier with the same fields for every model: identity, size, family & lineage, job, benchmarks & Court results, sampling controls, free limits, both paid paths, data policy, compute, and health. Every dossier field is a filter in this index, and the tag-match view shows every other model sharing a tag with the one you picked. Catalog presence is not free access. Where a data policy depends on the tier — Google and Mistral — the card prints the exact condition and its source instead of a bare trains / doesn’t-train flag. Every number carries its source; anything the report does not record reads Unknown — never an invented value.

The council rule

Same weight = same expert, regardless of provider. gpt-oss, Gemma 4, GLM-5.x, Kimi K2.x, Nemotron 3, Qwen, Llama, and DeepSeek each appear on 3–6 providers in this list. Distinct families buy independence; distinct weights buy the capability ladder. Endpoints do not.

How “free” is defined here: no credit card required for the recurring tier, unless a card says otherwise. “Free” in the report means three different things in practice — a real recurring free tier, a one-time or trial credit, and a catalog listing that 402s or paywalls on first call. Section C cards are the third kind: listed so you don’t count them.

—cards on this page — one per distinct weight, deduped across providers, plus the local route and the 5 trap cards
—distinct model families across the 82 cards — families buy independence, endpoints don’t
—cards flagged VERIFIED LIVE (live API / catalog pull)
—cards flagged UNVERIFIED (web-search only) — plus 5 “not free” trap cards (§C) = all 82 accounted for

⚖️ The Court’s ten core free weights — start here

The strongest no-card free core from the report (§D), deduped by weight. Tap any weight to open its full dossier; every other model in the index uses the exact same dossier fields, so any two dossiers compare tag-for-tag.

Container 1 — SpeedPay to go faster

The same model / same weight, where paying buys speed and throughput only — higher rate limits, priority queues, faster inference. No new capability. If a card says YES here, free and paid run the identical expert; paid just answers sooner.

Container 2 — VolumePay to do more

A free platform where paying buys more — more requests per day, bigger or additional models, more concurrency, extra features, or a production license. Kept separate from speed: a provider can sell one, the other, or both, and the cards score them independently.

Container 3 — DataData training

Every card is flagged Trains on your data, Does not train, Tier-dependent, or Unknown — sourced per provider, never guessed. Where the report states no policy, the card says Unknown. Starting points from your own records are shown with their caveat: Cloudflare’s zero-retention posture, Groq’s 30-day hold unless Zero Data Retention is flipped by the account owner, and Token Harbor free routes possibly retaining / using traffic (Token Harbor has no card — it is not in the verified report — so it stays Unknown until sourced). Two providers are tier-dependent, and their cards print the condition, not a bare flag (terms re-checked first-party on 2026-10-26): Google — on the unpaid tier Google uses your content and responses to improve its products and human reviewers may read them; on a billing-enabled project it states it does not. Mistral — free mode may use your input and output for training; pay-as-you-go keeps control with an opt-out toggle in the Admin panel.

Container 4 — ComputeWho pays for the GPU?

API access — the provider runs the compute and the free tier is the subsidy. Self-hosted — the weights are free but the compute is the bill (your hardware, or a rented GPU). A “free model” never hides a GPU bill on this page: self-hosted cards live in their own bucket and carry this container up front.

No cards match those filters. Loosen a filter or clear the search.

Council takeaways — from the report (§D)

  1. Strongest no-card free core, by role: reasoning — gpt-oss-120b (Groq / Cloudflare / Ollama), Llama 3.3 70B (Cloudflare FP8 / SambaNova), Nemotron 3 Super / Ultra (Ollama / OpenRouter / NVIDIA); code — Kimi K2.7-code, GLM-5.3, Qwen2.5-Coder-32b, Codestral (Mistral); fast — gpt-oss-20b, Llama 3.2 1b/3b, Gemma 4 26b, Liquid LFM 2.5 2.6b; guard — Llama Guard 3, gpt-oss-safeguard, prompt-guard (guards only, never council votes).
  2. Biggest free-volume routes: Mistral Experiment (no published cap, ~1 RPS estimated) and NVIDIA NIM (~40 RPM, prototyping-only license) have the most headroom; Groq (1K RPD / 200K TPD) and OpenRouter (50 RPD base) are the tightest. A council that fans out across providers survives any single cap.
  3. Needs a live check before anyone builds on it: every web-search ID in Section B, the Cloudflare paid-gate flags, the Google exact model IDs, and the Mistral console list. Free pools rotate monthly.
  4. Local counts differently: Rob’s orchestrator brief explicitly excludes local models — local counts for Maxx’s own council, not for Rob’s design (report §B8).

Tech research — capture hardware & the free audio stack

Folded in from the Tech Research thread (2026-10-26): Snap SPECS, the Omi wristband, and the local diarization work. Research notes, not model cards — verdicts and limits are stated exactly as recorded, not softened.

1 · Snap SPECS AR glasses PARKED — NOT BOUGHT

Snap’s first consumer AR glasses (Specs Inc.): pre-order at specs.com, shipping “later this fall” 2026 in US / UK / France, no exact date given. $2,195 base; $2,395 Connected Case bundle; $200 refundable deposit; 14-day return, 1-year limited warranty. Fully standalone (dual Snapdragon, no phone / puck / tether), see-through LCoS waveguide, 51-degree field of view, 132–136 g, 4 cameras, 6 MEMS microphones, hand tracking, spatial audio. Comparators: Meta Ray-Ban Display $799, regular Meta Ray-Bans ~$299–350 (no display), Apple Vision Pro $3,499.

  • Not a capture device: Snap says SPECS do not continuously record or save ambient audio, photo, or video, and have no facial recognition — an LED lights on intentional record only. This is a display computer, not a recorder; do not spec it into an always-on capture loop.
  • Battery: ~4 hours mixed use (~20 hours with the case) — not a full work day.
  • Other friction: hands-on reports call hand-pinch controls finicky and note heat and ear marks; Virtual Try-On requires an iPhone XS or later (an Android user cannot complete it); SPECS Intelligence previews iOS / Mac first.
  • Watch triggers: exact ship date, independent long-wear reviews, Android support, price drop, or Gen 2. No pre-order, no deposit paid.

2 · Omi wristband — diarization yes, named ID is Pro-only

The Omi Wrist Band is an accessory, not a separate device — the same Omi pendant snaps out of its necklace mount into the wristband. Diarization lives in the Omi system (app / backend), not the band hardware, so wrist vs. necklace changes capture position only.

  • Omi’s real-time transcription provides speaker identification, and its batch transcription has a diarize option attaching speaker labels to segments — format {text, speaker, start, end}, labels like SPEAKER_00.
  • The gate: speaker identification is Pro-only in Omi’s hosted service — the free plan shows it as not included; Pro ($29.99/month or $269.91/year) includes it. Free plan: unlimited on-phone transcription plus 300 cloud minutes/month.
  • Accuracy caveat: diarization alone gives anonymous labels, not names. Omi’s own docs warn diarization can split one voice into too many clusters, grouped as “Others”. Named identification needs a speech profile / embedding layer on top (see §4).
  • Omi is MIT open source and supports a custom backend URL / self-host — which is what makes the shim in §3 possible. Its former Deepgram diarization was retired 2026-08-27; current serving defaults are Modulate (velma-2), Soniox, and Parakeet.

3 · Omi custom-STT shim PROTOTYPE V1 — VERIFIED LOCALLY

An aiohttp WebSocket implementing Omi’s External Custom STT contract (binary audio frames, typically Opus 16 kHz, plus a JSON {"type": "CloseStream"} in; {"segments": [...]} back). It reuses the already-verified local pipeline — faster-whisper “base” + pyannote.audio 4.0.7, fully offline from the local Hugging Face cache. No cloud STT, no Omi Pro, no token at runtime.

  • Bronze tee (the point of the shim): on CloseStream, before replying to Omi, it writes a local bronze copy — bronze/<slug>.ogg (raw stream), bronze/<slug>.diarized.txt ([start-end] SPEAKER_xx: text), and bronze/<slug>.segments.json (the exact segments sent to Omi) — guaranteeing a local capture copy regardless of what the Omi app stores.
  • Verified 2026-10-26 with a simulated Omi app: a 15-second Ogg-Opus clip sent as 8 KB binary frames + CloseStream returned 1 segment in 47 seconds wall time — [2.86-14.34] SPEAKER_00: “Okay, so we've got a, we've got an audio recording right now…”. One bug found and fixed in that test: a boundary word with zero diarization overlap came out as SPEAKER_??; the shim now inherits the neighbouring speaker and re-merges, so “??” never reaches Omi. The server was stopped after the test; start it manually per session.
  • Batch, not real-time: segments return only after CloseStream, because pyannote needs the whole file and is slow on a 2-core, no-GPU VM (a 47-second clip previously took ~116 seconds for diarization alone). Fine for bronze capture; late for Omi’s live view.
  • Localhost only, no auth: a phone running the Omi app cannot reach 127.0.0.1 on this VM. Phone use needs a reachable endpoint (Tailscale / Fly / home server) + auth. Never expose v1 publicly.
  • Ogg-Opus only, proven: a real Omi app may send raw Opus packets without an Ogg container — that framing is untested until a real Omi device connects.
  • Anonymous labels only: SPEAKER_00/01, not names. Named ID is v2. Omi-app-side custom-STT wiring has not been tested from a real phone. The shim is not real-time, not phone-ready, and not name-identifying — yet.

4 · The free audio stack — transcription + diarization + named ID, no subscription

No single open model does all three jobs. The stack is three parts, all free / open:

  • Transcription: NVIDIA Parakeet TDT 0.6B v3 (word-level timestamps, for joining words to speaker turns) or faster-whisper / WhisperX (faster-whisper + pyannote, aligned so each word carries a speaker tag). Whisper alone does no diarization.
  • Diarization — NVIDIA Nemotron-3-Diarization: free/open, ~100M parameters, up to 8 speakers, overlapping speech native, real-time streaming buffers 0.32–30.4 seconds at 10 ms resolution, leads VoiceArena’s diarization bench at 14.72% DER, pairs with Parakeet. Cap: 8 speakers; labels are anonymous.
  • Diarization — pyannote.audio 3.1 / community-1: leading open diarization pipeline, ~11% DER on AMI, no hard speaker cap, heavier and slower on CPU. community-1 improves speaker counting/assignment and adds an “exclusive” mode that simplifies STT alignment.
  • Named identification (separate layer, required for names): a speaker embedding model — ECAPA-TDNN, WeSpeaker, or pyannote embeddings. Enrol a short sample per person, compare each diarized segment’s embedding to the enrolled profiles, map SPEAKER_xx to a name, and persist that profile. Shortcut that bundles diarization + centroids: paraspeakrs (Parakeet ASR + speakrs diarizer, no account or token, per-speaker centroid embeddings returned directly).
  • Summaries, tasks, and memories are not an audio model at all — that is the LLM layer on top of the transcript, which the existing stack already provides.