Shipped

Shippedhttps://shipped.anth.us/2026-10-21T00:00:00ZShippedAnnouncements and real-world deployments of AI, checked against their sources and scored: open-weights releases, decision models in the wild, and what is actually in use.REAX 1.1: a 27B decision model fine-tuned from pplx-deciderhttps://shipped.anth.us/decisions/decisionmodels-reax-1-1/2026-10-21T00:00:00Z2026-10-21T00:00:00Z

Apache-2.0 weights built on REAX 1, itself a fine-tune of pplx-decider-v1.2. The card reports 71.27 on its own SN92 scoring.

Decision4Rec releases Kev models that rerank recommender candidateshttps://shipped.anth.us/decisions/polignano-decision4rec-kev/2026-10-19T00:00:00Z2026-10-19T00:00:00Z

Apache-2.0 adapters from 0.8B to 27B that read a pool of 30 candidates and output a probability for each. Development results only.

AccuKnox Inc. releases Guard 1.1, an Apache-2.0 guardrail for agent inputs and prompt injectionhttps://shipped.anth.us/decisions/accuknox-guard-1-1/2026-10-16T00:00:00Z2026-10-16T00:00:00Z

A 4.5B guard fine-tuned from a Qwen3.5-4B base that returns a block probability. The card reports 0.802 on a 27,416-item safety set.

Maincode releases Matilda Decision v1, a 27B multimodal decision modelhttps://shipped.anth.us/decisions/maincode-matilda-decision-v1/2026-10-15T00:00:00Z2026-10-15T00:00:00Z

Apache-2.0 weights on Qwen3.8-27B. The maker's own run scores 65.38 on the public Jev Decision Index 0.3 suite.

decisionmodels releases s1-llm-auto-router, a 183 MB CPU router under MIThttps://shipped.anth.us/decisions/decisionmodels-s1-llm-auto-router/2026-10-15T00:00:00Z2026-10-15T00:00:00Z

A pruned mmBERT-small that answers seven routing questions in one CPU pass. The card reports 0.898 category accuracy on its held-out set.

Liquid AI releases LFM2.5-24B-A2B, a 24B on-device reasoning MoE with 2B activehttps://shipped.anth.us/open-weights/liquidai-lfm2-5-24b-a2b/2026-10-12T00:00:00Z2026-10-12T00:00:00Z

LFM2.5-24B-A2B is a 24B-parameter reasoning model with about 2.45B active, built for agents on-device, under the LFM1.0 licence.

StartLux releases Decision v1.1, from 0.8B to 27B, with non-commercial weightshttps://shipped.anth.us/decisions/startlux-decision-v1-1/2026-10-11T00:00:00Z2026-10-11T00:00:00Z

StartLux-Decision-v1.1 ships Instant and AutoThink variants. Weights are CC BY-NC 4.0; commercial use needs a separate licence.

AI Singapore releases Gemma-SEA-LION-Guard v2, MIT-licensed safety models for Southeast Asian languageshttps://shipped.anth.us/decisions/aisingapore-sea-lion-guard-v2/2026-10-11T00:00:00Z2026-10-11T00:00:00Z

A 31B decision model (plus E4B and E2B) for content moderation and action governance across eight languages. Scores AUPRC 93.5 on general safety.

Microsoft releases Decision-1, a 9B model that scores fixed answer optionshttps://shipped.anth.us/decisions/microsoft-decision-1/2026-10-09T00:00:00Z2026-10-09T00:00:00Z

Microsoft-Decision-1 returns a probability per answer option instead of prose. Hosted on Foundry and OpenRouter; no licence stated.

Amazon publishes Strands Decider 2B, a small decision model for agent workflowshttps://shipped.anth.us/decisions/amazon-strands-decider-2b/2026-10-09T00:00:00Z2026-10-09T00:00:00Z

A 2B LoRA on Qwen3.5-2B-Base, Apache-2.0, with a /v1/systemone server. The card reports 0.779 on 231 public JevBench tasks.

Reflection releases Beam, a 501B open-weights MoE under Apache 2.0https://shipped.anth.us/open-weights/reflection-beam-1-0-501b-a23b/2026-10-07T00:00:00Z2026-10-07T00:00:00Z

Beam is Reflection's first open model: 501B total and 23B active parameters, 1M-token context, under Apache 2.0.

OpenAI opens its Decisions API in public beta, built on GPT-6 Lunahttps://shipped.anth.us/decisions/openai-decisions-api/2026-10-06T00:00:00Z2026-10-06T00:00:00Z

A /v1/decisions endpoint returns typed predicate, choice and score answers. Public beta, $0.10 per million input tokens, GA promised in weeks.

Torchcast AI releases Torchcast Decision 27B, built on StartLux weightshttps://shipped.anth.us/decisions/torchcast-decision-27b/2026-10-03T00:00:00Z2026-10-03T00:00:00Z

A LoRA fine-tune of StartLux-Decision-27B that speaks the /v1/systemone API. Weights are CC BY-NC 4.0; card reports 65.00 on Index 0.2.1.

Aleph Alpha releases Kolibri 1, a 78B German and English reasoning MoE under Apache 2.0https://shipped.anth.us/open-weights/aleph-alpha-kolibri-1/2026-10-03T00:00:00Z2026-10-03T00:00:00Z

Kolibri 1 is a 78B mixture-of-experts reasoning model with 3.46B active parameters, tuned for German and English, under Apache 2.0.

SberAI releases FRIDA-Decisions, a Russian and English encoder for structured decisionshttps://shipped.anth.us/decisions/sberai-frida-decisions/2026-10-02T00:00:00Z2026-10-02T00:00:00Z

A 0.8B MIT-licensed encoder that picks, rates, answers yes/no or ranks in one pass. The card reports 0.893 on 735 razvilka items.

H2O.ai releases H2O Lightning 4B, an Apache-2.0 decision model with GGUF buildshttps://shipped.anth.us/decisions/h2oai-h2o-lightning-4b/2026-10-02T00:00:00Z2026-10-02T00:00:00Z

Built on Qwen3.5-4B, with GGUF files from 1.43 GB for llama.cpp. The card reports 88.74% on 231 public JevBench text questions.

Perplexity releases pplx-decider, an Apache-2.0 27B decision modelhttps://shipped.anth.us/decisions/perplexity-pplx-decider/2026-10-01T00:00:00Z2026-10-01T00:00:00Z

pplx-decider-v1 and v1.1 are open-weight 27B models that return probabilities for fixed answers. Built on Qwen3.8-27B.

Fish Audio releases Drama 3, a 35B mixture-of-experts TTS for multi-character sceneshttps://shipped.anth.us/open-weights/fish-audio-drama-3/2026-10-01T00:00:00Z2026-10-01T00:00:00Z

Drama 3 is a 35B MoE speech model (3B active) with up to 16 speaker tags and voice cloning; research licence.

vLLM Semantic Router team releases Decision 2.0 open decision modelshttps://shipped.anth.us/decisions/vllm-decision-2-0/2026-09-29T00:00:00Z2026-09-29T00:00:00Z

Decision 2.0 is an Apache-2.0 family from vllm-sr. The 27B Vega card reports 74.0 on JevArena. Lux is a 9B sibling.

An independent developer releases Jev-Style 2B Decision v3, a 2B model with a free hosted APIhttps://shipped.anth.us/decisions/chaoliangunsw-jev-style-2b-decision-v3/2026-09-27T00:00:00Z2026-09-27T00:00:00Z

Apache-2.0 fine-tune of Qwen3.5-2B. The card reports 73.6% on public JevBench items against Jev 1.13 at 86.6%.

Shanghai AI Laboratory releases Intern-Decision, multimodal decision models from 0.8B to 4Bhttps://shipped.anth.us/decisions/internlm-intern-decision/2026-09-26T00:00:00Z2026-09-26T00:00:00Z

Intern-Decision-4B is fine-tuned from Qwen3.5-4B and answers up to 16 questions per pass, with image input. Apache-2.0.

Xiaomi MiMo releases MiMo-V2.6-Pro-RL, a 1.02T omnimodal MoE under MIThttps://shipped.anth.us/open-weights/xiaomimimo-mimo-v2-6-pro-rl/2026-09-21T00:00:00Z2026-09-21T00:00:00Z

MiMo-V2.6-Pro-RL is the flagship of the MiMo-V2.6 series: 1.02T total, 42B active, 1M context, text, image, video and audio in.

distil labs releases an invoice-decision model that approves or holds vendor invoiceshttps://shipped.anth.us/decisions/distil-labs-qwen3-5-4b-invoice-decision/2026-09-21T00:00:00Z2026-09-21T00:00:00Z

A 4B Qwen3.5 fine-tune returning one of five decisions. The card reports 98 of 100 correct on synthetic test invoices.

JetBrains releases Mellum2.1 Thinking, a 12B coding MoE with 2.5B active, under Apache 2.0https://shipped.anth.us/open-weights/jetbrains-mellum2-1-12b-a2-5b-thinking/2026-09-20T00:00:00Z2026-09-20T00:00:00Z

Mellum2.1 Thinking is a 12B mixture-of-experts reasoning model with 2.5B active, retrained mostly with reinforcement learning on real repositories.

Jared Palmer releases Kev-9B, an Apache-2.0 LoRA decision model on Qwen3.5-9Bhttps://shipped.anth.us/decisions/jaredpalmer-kev-9b/2026-09-20T00:00:00Z2026-09-20T00:00:00Z

An independent adapter that implements TypeSafe's System One API. The card reports 0.852 on a locked out-of-domain test.

Convai Innovations releases laya-typed-decisions, a 0.4B ModernBERT decision modelhttps://shipped.anth.us/decisions/convai-laya-typed-decisions/2026-09-18T00:00:00Z2026-09-18T00:00:00Z

An Apache-2.0 encoder fine-tuned on four synthetic workflows. The card reports 0.766 accuracy and warns its probabilities are over-confident.

TypeSafe AI launches Jev, a hosted model for typed software decisionshttps://shipped.anth.us/decisions/typesafe-jev/2026-09-15T00:00:00Z2026-09-15T00:00:00Z

Jev takes state and typed questions and returns probabilities. Early access via API, $0.042 per million input tokens; weights not stated.

Yandex releases AliceAI-Foundation-80B-A3B-Base, an Apache 2.0 base MoE for Russian and Englishhttps://shipped.anth.us/open-weights/yandex-aliceai-foundation-80b-a3b-base/2026-09-12T00:00:00Z2026-09-12T00:00:00Z

AliceAI-Foundation-80B-A3B-Base is an 80B base model with 3B active parameters and 262,144-token context, trained from scratch.

DeepSeek releases V4.1-Flash, a multimodal MoE with a much smaller KV cache, under MIThttps://shipped.anth.us/open-weights/deepseek-ai-deepseek-v4-1-flash/2026-09-10T00:00:00Z2026-09-10T00:00:00Z

DeepSeek-V4.1-Flash is a 552B-backbone multimodal MoE with 1M-token context and a 890-byte-per-token KV cache, under MIT.

OpenBMB releases MiniCPM5-2B, a dense 2B model with open training data, under Apache 2.0https://shipped.anth.us/open-weights/openbmb-minicpm5-2b/2026-09-06T00:00:00Z2026-09-06T00:00:00Z

MiniCPM5-2B is a dense 2B on-device model for coding, tool use and long context, released with its UltraData training sets.