Judgment becomes a model class
Sarvam open-sources a 105B decision model — probability for every option in one forward pass, Apache-2.0, eleven Indic languages — while WeVisCut’s paper matches Claude Opus 5 at 33× fewer tokens on weights that still ship no license, and December 1 becomes a four-event migration day.
Sarvam-Prism (Apache-2.0, Nov 2) fine-tunes Sarvam-105B into a decision model — per-option probabilities in a single forward pass, 44 ms median on one B200, 131,072-token questions, eleven Indic languages plus English, no card-published accuracy evals yet; WeVisCut’s owed read lands a 100-task benchmark with Opus-5 parity at ~33x fewer generated tokens per output minute, and its weight-license read confirms no LICENSE file exists; Gemini 2.5 Pro and Flash deprecate with same-day December 1 auto-routes, stacking Argon’s price expiry into a four-event day. Issue 19, coverage window 2–4 November 2026.
Week of November 2–4, 2026 · Issue 19
The Week in One Sentence
A quiet window for launches turned loud on exactly the layers this publication tracks anyway: India shipped the decision-model tier’s first open heavyweight — Sarvam-Prism, a 105B Apache-2.0 model that returns a probability for every option in one forward pass — while two owed reads discharged with substance (WeVisCut’s paper claiming Claude-Opus-5 parity at thirty-fold fewer tokens, on weights whose license file still does not exist), Google stacked two same-day auto-route deprecations onto December 1, and Anthropic moved conversation state server-side and switched on text watermarks.
The Big Story: Sarvam-Prism — judgment as a first-class model
What happened. (India.) November 2, per the card: Sarvam released Prism — a decision model fine-tuned from Sarvam-105B, Apache-2.0, ~105B parameters in a custom FP8 layout (a prism.py loader, not AutoModelForCausalLM). The interface is the whole idea: given a context and questions with fixed answer options, Prism returns a probability for every option in a single forward pass — three question types (choice with per-key probabilities, yes/no, scored levels with an expected score) — and, in the card’s own words, “it does not generate text, so the answer is always one of your options.” Context runs to 131,072 tokens per question; languages cover English plus ten Indic (Hindi, Bengali, Tamil, Telugu, Marathi, Gujarati, Kannada, Malayalam, Oriya, Panjabi); the card reports median 44 ms per request (p95 420 ms) on a single NVIDIA B200, needing ~120 GB of FP8-capable VRAM — or the hosted API.
What the evidence says — including what isn’t there. The speed numbers are card-published and the architecture is fully specified (MoE backbone, 128 routed experts per layer as grouped FP8 matmuls, a small decision head reading hidden states at each option’s end and at an answer marker). What the card does not carry: any accuracy benchmark, any calibration number, and any limitations section — evaluation is deferred to the release blog and a promised “public Decision Index benchmark,” and the training data is not stated on the card. We label that precisely: the mechanism is open and inspectable, the quality of its judgments is currently a claim with no public table behind it.
Why this matters. The decision-model tier has assembled inside a month: OpenAI’s Decisions API priced the small fast judge at $0.10/M (Issue 11), Liquid’s d1 shipped typed answers with calibrated probabilities at 2B scale (Issue 18), and now a 105B open-weights entry tuned for Indic languages. Every agent stack has points where it needs cheap, calibrated judgment — routing, gating, classification, escalation — and is today burning generated tokens to get it. A model whose entire output is a distribution over your own options is that judgment as a primitive: auditable, cacheable, and — in Prism’s case — hostable. The open question is the one the card skips: whether the probabilities are calibrated (do 70% claims happen 70% of the time?), which is the difference between a decision model and a confident one.
What I would test. The metric the card doesn’t carry: run Prism as a router on two weeks of your own traffic — tickets to queues, requests to tiers — against two arms: your current classifier, and the base model prompted to choose. Score all three on calibration (reliability diagram, expected calibration error, Brier score) and not just accuracy, and log cost per thousand decisions. If an open 105B matches your classifier with honest error bars, the routing layer of your agent stack just got cheaper and inspectable at once.
Elsewhere in AI Engineering
WeVisCut: the paper lands, the license doesn’t (China) — Worth Reading
What happened. (China. The owed read from Issue 18, discharged.) The paper (arXiv:2610.38352, October 28; read at abstract depth this cycle — full text still queued) (Mu et al. 2026) describes the editing agent’s real novelty: finished videos don’t reveal editing decisions, so the team built a four-stage data pipeline of editing inversion and verified replay, applied to 3,750 finished videos to synthesize 130K+ training examples, training a single 9B policy inside an agent harness that keeps shared production state — tool feedback guides later decisions and repairs. On the paper’s 100-task benchmark (five metrics), WeVisCut “substantially outperforms its base model,” matches Claude Opus 5 while using nearly 33× fewer generated tokens per successful output minute1, and beats task-matched native editors. And the second debt: the weight-license read (November 6) finds no LICENSE file in the repository at all — the code stays Apache-2.0 on GitHub, the weights stay under separate, unstated terms.
Why this matters. The 33× number is the token-economics thesis with a number attached — harness-native efficiency, not model scale, buying the parity — and it lands in the same fortnight as the orchestration-beats-model ablations this publication keeps citing. But the deployment arithmetic is now stark: a state-of-the-art open agent whose weights no one may legally use. Capability shipped; terms didn’t. The flag renews weekly.
Hy-Nomas’s technical report reaches arXiv — the numbers still haven’t (China) — Worth Watching
What happened. (China. In-window event: the card cited the paper on November 4; the paper itself is dated October 27 — disclosed.) The technical report (Chen et al. 2026) confirms what the card wouldn’t: an 8B diffusion backbone (Issue 16 flagged the missing parameter count), structured video captions with action-aware causal attention — descriptive context persists across segments while actions apply only to their time windows — a vision-language Prompt Enhancer, and a curriculum that includes training on model-generated histories before diffusion distillation. The evaluation, in the abstract, is human-only: “superior performance over existing world models.” No quantitative tables yet — again. Weights, code, the benchmark suite, and an adaptation workflow are released; downloads run 81.3k (from 53.4k at the November 1 read).
Why this matters. The interactive-world-model line is compounding adoption fast and publishing evidence slowly; the suite ships with the repo, so the community now owns the instrument that could fill the gap. First external numbers are the watch.
Anthropic moved state server-side — and switched on text watermarks (US) — Worth Reading
What happened. (US. Both entries in-window.) November 2: the Messages API now offers server-side conversation threads (beta header message-threads-2026-08-12) — smaller requests, faster time-to-first-token, threads expiring after five idle minutes. November 3: text from Claude Mythos 5, Opus 4.7, and Sonnet 4.6 now carries Anthropic’s text watermark — with the caveat, stated on the note itself, that Sonnet 4.6 output may lack it under non-default sampling.
Why this matters. Threads are the externalized-state thesis arriving in API form: conversation as a server-side resource the client references, not a payload it re-sends — the same state-outlives-the-turn move as memory banks (Issue 14) and cached generation histories (Issue 16), now with a five-minute lease. The watermark is the provenance line this publication has tracked since SynthID — model output now carries vendor provenance by default at two of the three major APIs. Both are quiet changes with compliance-shaped consequences: audit trails get easier, and “which model wrote this” becomes a property of the text itself.
December 1 became a four-event day (US) — Worth Watching
What happened. (US. Labelled update on the migration calendar.) November 2 changelog entries: Gemini 2.5 Pro deprecated — requests auto-route to Gemini 3.8 Flash from December 1; Gemini 2.5 Flash deprecated — auto-routes to Gemini 3.5 Flash-Lite the same day. That stacks onto the standing clock: Argon’s intro rates end December 1 (rate-card lineage from the October 14 read; the changelog page itself doesn’t restate it), making it a two-migration, one-price-change day — with Gemini 3.7 Flash’s intro price running through December 31 as the next marker on the same page. Re-verified this cycle: Sonnet 4.5 and Opus 4.8 fast mode both retire November 30 (Anthropic), and grok-voice-2.0 remains Deprecated with no shutdown date stated anywhere (rate card re-read) — the undated-deprecation carry, older than every dated clock around it. And the voice-stack tidbit: Gemini 3.8 TTS gained voice remixing November 4 — prebuilt or designed voices modified by natural-language prompts into stored voices.
Why this matters. Auto-routing is the polite retirement: nothing breaks on December 1, but cost and quality silently change — 2.5 Pro traffic wakes up on a Flash-class bill and a Flash-class ceiling. If you have dated assertions about which Gemini answers your traffic, December 1 is the day to re-measure them.
The adoption-without-terms ticker (China; US carry) — Worth Noting
What happened. (Labelled updates, all re-read November 6.) DeepInfer: 25,516 stars, 3,124 forks, 247 open issues+PRs — and its sixth commit (November 3, in-window) is “docs: update V4.1 Flash Q&A to v0.2,” the first content motion since release, docs-scoped; still no published throughput numbers anywhere. SAS: twelfth license check — still no LICENSE — with reproduction priced at 192–400 H20 GPU-hours, a single-node job separated from legitimate retraining by one missing file. EmbeddingGemma 2 (US carry): 482,581 downloads — nearly double in five days (from ~256K on November 1) — and the tree re-read still shows no LICENSE or NOTICE, fifteen root files, terms living in blog prose.
Why this matters. Growth against absence is now measurable weekly, and this week it compounded on all three artifacts at once. When adoption outruns artifacts by these margins, the license file — not the model — becomes the decision-relevant fact. Read it, or wait.
Research Radar
No fresh in-window paper met the read-beyond-abstract bar — the window’s research substance was two owed reads discharged: WeVisCut above (abstract-depth; full text queued) and the Hy-Nomas technical report (abstract-depth). Sunday’s queue already holds paper-shaped leads (UniCom, November 5).
Repo Radar
sarvamai/prism (India)
- What’s actually in it: the ~105B weights in a three-precision FP8 layout, the
prism.pyloader, and a fully specified card (architecture, deployment, speed). Inspection depth: card in full; not executed. - What you can run: decision endpoints — classification, routing, gating — on one FP8-capable GPU with ~120 GB, or via Sarvam’s hosted API.
- What’s missing: accuracy and calibration evals (deferred to the blog and the promised Decision Index), a limitations section, and any standard-tooling support (custom loader).
- Why open it: the first open 105B decision model — and the calibration test it implicitly invites is the experiment below.
nvidia/DeepSeek-V4.1-Flash-1017-NVFP4 (US)
- What’s actually in it: NVIDIA’s own NVFP4 build of DeepSeek’s -1017 refresh — created October 30 (inside Issue 18’s window, uncaught by that run; caught here), updated in-window November 4. MIT, 6k downloads at read.
- What you can run: the biggest same-endpoint agent-suite jump this publication has ledgered (Terminal-Bench 4.0 31.2 → 42.3, vendor-run), in a 4-bit floating-point format tuned for NVIDIA silicon.
- What’s missing: NVIDIA-published benchmark deltas for the quantized build; independent evals of either precision.
- Why open it: the serving-cost artifact of the quarter’s best free-lunch refresh — now in the format its host hardware likes best.
My Read on the Week
(Synthesis and interpretation — not reported fact. Cross-item pass over all three desks’ items; every figure traces to a ledger row.)
1. Judgment is becoming a model class — and it is the agent stack’s newest organ. The decision tier assembled in under a month: an API priced at $0.10/M (Issue 11), a 2B edge family (Issue 18), and now a 105B open-weights entry covering eleven languages. The pattern rhymes with every organ this publication has watched split off — parsing, retrieval, memory-writing, now judgment: a general capability becomes a specialist artifact with a narrower output and a sharper contract. Prism’s contract — your options, their probabilities, one pass, no prose — is exactly what routers and gates need, and exactly what prompted general models do badly and expensively. The tier’s missing instrument is calibration, and whoever publishes the first reliability table owns the category’s evidence base.
2. The capability-terms gap is now the industry’s most reliable lag indicator. WeVisCut’s 33× efficiency result and its absent LICENSE file shipped in the same artifact; EmbeddingGemma 2 nearly doubled downloads in five days on prose-only terms; DeepInfer grew 800 stars in a week against zero published numbers; SAS’s reproduction is priced and blocked by one missing file. Four ecosystems, one shape: adoption first, evidence second, terms last. The buyer’s discipline that follows is already in this publication’s rules — no artifact decision without the license file read — and this week is the best argument yet for extending it: the more impressive the artifact, the more likely its terms are the missing part.
3. State keeps moving out of the request — in both directions across the API-office boundary. Anthropic’s threads put conversation state on the server with a five-minute lease; Hy-Nomas trains on model-generated cached histories; Prism reads a 131k-token context per question but answers in a single pass. The same design question — where does state live, who owns its lifetime — is being answered at the API layer, the training layer, and the architecture layer in the same month. The prize is the same each time: don’t re-send, don’t re-generate, don’t re-judge.
What I’d Watch Next
- Sunday’s already-loaded queue (November 5–6, ledgered as leads): Kimi-K3.1-Pro — weights touched November 5 and hosted on Sarvam’s Token Factory the next day at ₹250/₹25/₹1,350 with kimi-k3 discontinued — plus Qwen3.8-27B, Liquid’s LFM3 quant wave, NVIDIA’s MetricTrail and Nemotron-3.5-Parse, and Molmo2.
- Bulbul V3’s Voice-Agents EOL executes November 10 — six days from this issue’s date, re-verified verbatim from the changelog primary; TTS-API users unaffected.
- The migration calendar and the two open evidence debts — November 30 (Sonnet 4.5, Opus 4.8 fast mode), December 1 (2.5 Pro/Flash auto-routes, Argon intro end); WeVisCut’s weight terms; Prism’s Decision Index evals.
One Thing I’d Experiment With This Week
Build the calibration harness — the table Prism’s card doesn’t carry. You can run the experiment the release deferred, on your own traffic, in an afternoon.
- Setup: take two weeks of a real decision stream you already label — support tickets to queues, requests to severity tiers, emails to act/defer/drop. Format each as Prism’s schema (context + fixed options), and mirror the same cases into two arms: your current classifier, and a general model prompted to pick (temperature 0, options in the prompt).
- Conditions: score all three arms against your ground-truth labels, but bin every prediction by confidence decile first — the reliability diagram is the point, not the accuracy line.
- Measure: expected calibration error and Brier score per arm; accuracy at the operating threshold you actually use; cost per thousand decisions (Prism’s 44 ms median is per request — multiply honestly). Decision output: the first published calibration table for an open 105B decision model — and if Prism’s 70% bins hold at 70%, the routing layer of your agent stack has a new default; if they don’t, you’ve caught a confident model before it made a confident mistake, which is the entire reason this model class exists.
All vendor-reported figures are labeled as such; if independent evidence doesn’t exist yet, that’s stated rather than papered over. This issue’s desk ledgers live under research/2026-11-04/.
References
Footnotes
WeVisCut’s comparators (Claude Opus 5, its Qwen3.5-9B base) and Prism’s lineage (Sarvam-105B) are different generations from the comparator sets of Issues 11–18 — cross-table score comparisons across comparator vintages remain not apples-to-apples (EDITORIAL.md §6).↩︎
Citation
@online{dimsum_editorial2026,
author = {dimsum editorial, the},
title = {Judgment Becomes a Model Class},
date = {2026-11-04},
url = {https://news.llm101.in/issues/2026-11-04/},
langid = {en}
}