# A small fast cousin for sub-tasks beside a main model

> A small fast cousin for sub-tasks beside a main model: 8 companies fully do it, 18 with a limit, 90 carry the tag. Every listing in Open-weight models with the reason and a source.

- Canonical: https://www.anchorterminal.com/niches/open-weights.sub-task-cousin/
- Markdown: https://www.anchorterminal.com/niches/open-weights.sub-task-cousin/index.md (~5,150 tokens)
- Slim: https://www.anchorterminal.com/niches/open-weights.sub-task-cousin/index.min.md (~1,230 tokens, same facts, less prose, for token-sensitive contexts)
- JSON: https://www.anchorterminal.com/niches/open-weights.sub-task-cousin/index.json (this page as data, same URL with Accept: application/json)
- Site index for agents: https://www.anchorterminal.com/llms.txt (the key docs and every listing in one file, at most 16 MB, about 4 million tokens: https://www.anchorterminal.com/llms-full.txt)
- API: https://www.anchorterminal.com/api/v1/index.json
- MCP: https://www.anchorterminal.com/mcp (the whole catalogue as tools, no key)
- DNS: dig +short TXT <slug>.grade.anchorterminal.com (the grade as a DNS TXT record, for a sandbox with no HTTP where names still resolve; also top., best.<category>. and <a>.vs.<b>. under the zone, and the same under grade.letme.dev)
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)
- Updated: 2026-11-15

Four or more companies' docs show a full fit.

| Companies, full fit | With a limit | Carry the tag | Share of the category's work | First-hand accounts |
| --- | --- | --- | --- | --- |
| 8 | 18 | 90 | 8 | 7 |

## What counts as done

Two models fit co-resident within the machine's memory, the cousin runs with thinking off, and the main model is not reloaded between calls.

Measured, for our own tests: the cousin answers a routing or summary prompt correctly in under 3 seconds.

How agents ask for it: a cheap model for title generation and routing; something small for the harness's utility calls; a planner and a coder that both fit in memory; a second-opinion reviewer model

## Does the job

The docs show it meets the test above, and nothing we read says otherwise.

- **Gemma 4 12B IT (Google LLC (RDAP registrant of gemma.ai as recorded for the first batch's Gemma 4 listings; the card names Google DeepMind as author))**: Thinking is off by default, the Google Q4_0 GGUF is 7 GB, and Ollama's gemma4:12b (7.7 GB) carries the tools and thinking badges, so it sits beside a main model.
- **Gemma 4 E2B IT (Google LLC (RDAP registrant of gemma.ai as recorded for the first batch's Gemma 4 listings; the card names Google DeepMind as author))**: 3.35 GB Q4_0 from Google, thinking off by default, Ollama gemma4:e2b (4.6 GB) with tools and thinking badges; built for precisely this spare-memory role.
- **Gemma 4 E4B IT (Google LLC (RDAP registrant of gemma.ai as recorded for the first batch's Gemma 4 listings; the card names Google DeepMind as author))**: 5.15 GB Q4_0 from Google, thinking off by default, Ollama gemma4:e4b (6.6 GB) with tools and thinking badges, and acc-45's agent task passed.
- **Granite 4.2 3B (International Business Machines Corporation (per IBM's naming guidance, as read for granite-guardian))**: acc-26 found Granite 4 routed 33 voice prompts at 100 percent in 0.54 seconds on a 16 GB GPU; IBM ships GGUFs from 1.5 GB, Ollama has granite4.2:3b (2.2 GB) and enable_thinking false is documented.
- **Ling-3.0-tiny (Ant Intelligence (Hang Zhou) Technology Co., Ltd (named as provider in the training-content summary); inclusionAI, Ant Group in the site footer of inclusion-ai.org)**: 4.8 GB Q4_K_M co-resides easily, enable_thinking false switches reasoning off per request, and an account measured 40 tokens per second on a CPU and over 200 on an RTX 4070.
- **NVIDIA Nemotron 3.5 Bolt E4B 4GB (NVIDIA Corporation)**: 6.2 GB Q4_K_M from NVIDIA, enable_thinking false per request, a Brevity Benchmark row and an Ollama e4b tag of 7.2 GB to 7.9 GB with the tools badge.
- **Nemotron 3 Nano 4B (NVIDIA Corporation)**: 2.8 GB Ollama tag and NVIDIA Q4_K_M GGUF, enable_thinking false per request, so it sits beside a main model in under 3 GB; English only.
- **Qwen3-4B-Instruct-2507 (Alibaba Cloud (named in the licence's copyright line))**: Never thinks, 2.5 GB Ollama 4b-instruct tag with the tools badge, 262K context; a routing or summary call needs no switch.
- **Qwen3.5-0.8B (Alibaba Cloud (named in the licence's copyright line))**: Thinking is off by default; the card warns this size falls into thinking loops if you turn it on. Ollama qwen3.5:0.8b is 1.0 GB with the tools badge, so it sits beside a main model.
- **Qwen3.5-2B (Alibaba Cloud (named in the licence's copyright line))**: Thinking is off by default; the card warns this size falls into thinking loops if you turn it on. Ollama qwen3.5:2b is 2.7 GB with the tools badge, so it sits beside a main model.
- **Qwen3.5-4B (Alibaba Cloud (named in the licence's copyright line))**: Thinking switched off per request with enable_thinking false, Ollama qwen3.5:4b at about 3 GB with the tools badge, 2 GB of KV at 64K.
- **Qwen3-Coder-30B-A3B-Instruct (Alibaba Cloud (named in the licence's copyright line))**: No thinking mode, 19 GB Ollama tag with the tools badge; acc-51 kept it resident beside a 32B planner at about 40 GB total.
- **Qwen3-VL-2B-Instruct (Alibaba Cloud (per the first batch's RDAP check of qwen.ai; this model's files name no copyright holder))**: No thinking mode, publisher Q4_K_M GGUF of 1.1 GB, Ollama qwen3-vl:2b-instruct at 1.9 GB with the tools badge.
- **Qwen3-VL-4B-Instruct (Alibaba Cloud (per the first batch's RDAP check of qwen.ai; this model's files name no copyright holder))**: No thinking mode, publisher Q4_K_M GGUF of 2.5 GB, Ollama qwen3-vl:4b-instruct at 3.3 GB with the tools badge.
- **Qwen3-VL-8B-Instruct (Alibaba Cloud (per the pilot's RDAP check of qwen.ai; this model's files name no copyright holder))**: No thinking mode, publisher Q4_K_M GGUF of 5.0 GB, Ollama qwen3-vl:8b-instruct at 6.2 GB with the tools badge.
- **Spark-X2.6-4B (XHToken (the licence's copyright line); the Hugging Face organisation's name is SparkLLM)**: enable_thinking false or reasoning_effort low per request, publisher Q4_K_M GGUF of 3.5 GB, Ollama SparkLLM/Spark-X2.6-4B entry.

## Does the job, with a limit

It does the job, but something in the docs or a user's account will get in an agent's way. The limit is named in each row.

- **Agents-A1-4B (Shanghai Artificial Intelligence Laboratory (the InternAgent team there, per the card's recommended system prompt and the template's default prompt); the GitHub LICENSE leaves the copyright line as the Apache template)**: 2.7 GB Q4_K_M build from the publisher and enable_thinking false in the template, but thinking is on by default and the card never documents the switch; no Ollama entry.
- **Apodex-1.1-mini (Apodex (the site's copyright line reads 2026 Apodex))**: A 21.7 GB Q4_K_M GGUF fits 25 GB of spare memory and enable_thinking false is in the template, but thinking is on by default, the card does not document the switch and there is no Ollama entry.
- **BigBang-v1 (The Endless Frontier (endless-frontier))**: No publisher GGUF; bartowski's Q4_K_M is 21.9 GB and the template's enable_thinking switch is undocumented on the card. Thinking defaults on.
- **Darwin-36B-Opus (VIDRAFT Inc. (named on vidraft.net); the card's citation names FINAL-Bench and VIDRAFT_LAB)**: Publisher GGUFs down to IQ1_M (8.2 GB) exist and enable_thinking false is in the template, but the card presents it as an always-thinking reasoning model and never names the switch.
- **Gemma 4 26B A4B IT (Google LLC (RDAP registrant of gemma.ai; the card names Google DeepMind as author))**: Thinking off by default and an Ollama tag with tools, but the Q4_0 build is 14.4 GB (16 GB on Ollama), so it needs the top of the 2 to 25 GB spare band.
- **Gemma 4 31B IT (Google LLC (the registrant of gemma.ai; the card names Google DeepMind as author))**: Thinking off by default and an Ollama tag with tools, but 17.6 to 19 GB at Q4 uses most of a 25 GB spare band; a 31B dense is slow for a 3 second routing reply.
- **GLM-4.7-Flash (JINGSHENG HENGXING TECHNOLOGY PTE.LTD (operator of Z.ai per its terms of use, and registrant of z.ai))**: enable_thinking false is in the template and Unsloth says 24 GB of memory at 4-bit, but 17.5 to 19 GB of weights takes most of a 25 GB spare band.
- **gpt-oss-20b (OpenAI (named in USAGE_POLICY and the model card; the LICENSE file's copyright line is the unfilled Apache template))**: 12 to 14 GB fits a 25 GB spare band and reasoning low is cheap, but thinking cannot be switched off, so a routing reply still carries an analysis channel.
- **Laguna XS 2.1 (Poolside, Inc. (named in Poolside's Terms of Use))**: Thinking is off by default and 3B are active, but the smallest build is 20 GB, at the top of the spare-memory range.
- **LFM2.5-2.6B (Liquid AI, Inc. (the licensor named in the LFM Open License))**: 1.6 GB at Q4_0 and an Ollama entry with tools, but the card says it always thinks and the template forces a think tag, so routing calls carry reasoning time.
- **MiMo-V2.6 Distill Qwen 9B (Xiaomi (the card and report name Xiaomi MiMo; no LICENSE file names a copyright holder))**: 6.6 GB and enable_thinking false empties the think block, but it is an SFT checkpoint meant for RL and non-stop generation is reported.
- **MiniCPM-o 4.5 (OpenBMB (named as the MiniCPM-o team on the technical report; the GitHub LICENSE is the unfilled Apache template))**: Q4_0 is 4.8 GB and the template empties the think block on enable_thinking false, but tool calling is undocumented and the Ollama build carries only a vision badge.
- **MiniCPM5-2B (OpenBMB (named in the GitHub LICENSE's copyright line))**: 1.56 GB Q4_K_M and an Ollama entry, but every answer opens with reasoning; the template accepts enable_thinking false while the card documents only Think mode, and discussions report loops.
- **Muse Glimmer 30B (Meta (the verified Hugging Face organisation is "Meta Inc.", the card names Meta Superintelligence Lab, and the licence carries no copyright line))**: The 17 GB build fits a 25 GB spare budget, but reasoning_strength bottoms out at low with no off value, so a routing call still thinks.
- **Nanbeige4.3-33B (BOSS Zhipin (Kanzhun Limited), unconfirmed)**: enable_thinking false turns reasoning off, but the smallest usable 4-bit build is about 19 GB, most of a 25 GB spare budget.
- **Nemotron 3.5 Lightning 30B A3B (NVIDIA Corporation)**: enable_thinking false turns reasoning off and 3B active is fast, but the smallest build is 18.9 GB, most of a 25 GB spare budget.
- **Nemotron 3 Nano 30B A3B (NVIDIA Corporation)**: enable_thinking false and a reasoning_budget example on the card, 3.5B active, but 18 GB is most of a 25 GB spare budget.
- **Nemotron 3 Nano Omni 30B A3B Reasoning (NVIDIA Corporation)**: enable_thinking false gives an instruct mode, but the smallest GGUF is 18.5 GB, most of a 25 GB spare budget.
- **occamy-1.0 (Accio (Accio-Lab))**: Thinking switches off through the Qwen3.6 template, but the smallest GGUF is 11.7 GB (IQ2_M) and a 4-bit build about 19 GB, most of a 25 GB spare budget.
- **Ornith-1.5-35B-A3B (Ornith)**: 3B active and enable_thinking false in the template, but the card never mentions the switch and the smallest build is 21.7 GB.
- **Ornith-1.5-9B (Ornith)**: 5.8 GB and enable_thinking false in the template, but the card does not document the switch and the model is a long thinker.
- **Qwen3-30B-A3B-Instruct-2507 (Alibaba Cloud (named in the licence's copyright line))**: No thinking block at all, so routing calls answer at once, but the 4-bit build is 19 GB, most of a 25 GB spare budget.
- **Qwen3.5-9B (Alibaba Cloud (named in the licence's copyright line))**: Thinking off per request and Ollama qwen3.5:9b at 6.6 GB, but acc-26 measured it over eight times slower than Granite on routing, past the 3 second bar on a 16 GB GPU.
- **Qwen3.6-35B-A3B (Alibaba Cloud (named in the licence's copyright line))**: acc-36 maps a non-thinking variant to title generation beside the main model, but at 22 GB UD-Q4_K_XL it takes most of a 25 GB spare budget.
- **Qwen3-VL-30B-A3B-Instruct (Alibaba Cloud (per the pilot's RDAP check of qwen.ai; this model's files name no copyright holder))**: No thinking and 3B active, but 18.6 GB at Q4 takes most of a 25 GB spare budget.
- **SmolLM3-3B (Hugging Face, Inc.)**: enable_thinking false per request and 1.9 GB Q4 GGUFs from ggml-org and Unsloth, but Hugging Face ships no GGUF and Ollama has no library entry.

- not this job: Apertus v1.5 8B (ETH Zurich and EPFL (the Swiss National AI Institute, named in the Acceptable Use Policy)): No publisher GGUF, no Ollama entry and no stock runtime loads the architecture; a cousin cannot be loaded beside a main model.
- not this job: Carbon-X (Hugging Face (the HuggingFaceBio organisation, named Hugging Face Biology Research on the Hub)): Routing or summary prompts are outside its DNA training; the card's agent section makes it a tool an agent calls, not a cousin that routes.
- not this job: Cosmos-Reason2-2B (NVIDIA Corporation): NVIDIA ships no GGUF and no thinking switch beyond a system-prompt instruction; a routing cousin is not what it is for.
- not this job: DeepSeek-V3.2 (DeepSeek (the MIT licence names DeepSeek, and DeepSeek's platform terms name Hangzhou DeepSeek Artificial Intelligence Co., Ltd.)): No small build exists; the smallest file is 161 GB.
- not this job: DeepSeek V4.1 Flash (DeepSeek (the MIT licence names DeepSeek, and DeepSeek's platform terms name Hangzhou DeepSeek Artificial Intelligence Co., Ltd.)): No small build; 233 GB minimum.
- not this job: DeepSeek-V4-Flash-0731 (DeepSeek (the MIT licence names DeepSeek, and DeepSeek's platform terms name Hangzhou DeepSeek Artificial Intelligence Co., Ltd.)): Smallest build about 103 GB; it is the main model, not a cousin.
- not this job: DiffusionGemma 26B A4B IT (Google LLC (RDAP registrant of gemma.ai; the card names Google DeepMind as author)): No Google quantised build, no Ollama entry, and the only local GGUF needs a llama.cpp branch; output arrives one 256-token canvas at a time.
- not this job: GLM-4.5-Air (JINGSHENG HENGXING TECHNOLOGY PTE.LTD (operator of Z.ai per its terms of use, and registrant of z.ai, per the first batch's check on 21 October 2026)): The smallest build anywhere is 38 GB (UD-TQ1_0); it is a main model, not a cousin.
- not this job: GLM-5.3 (JINGSHENG HENGXING TECHNOLOGY PTE.LTD (operator of Z.ai per its terms of use, and registrant of z.ai). The licence names Z.AI): No small build.
- not this job: GLM-5.3-Flash (Z.AI Co., Ltd (named in the licence's copyright line)): Smallest build 93 GB.
- not this job: gpt-oss-120b (OpenAI (named as publisher on the card; the openai-api listing's terms name OpenAI OpCo, LLC)): 63 GB; it is the main model.
- not this job: Hy3 (Tencent (named in the LICENSE file's copyright line)): Smallest build 89 GB.
- not this job: Inkling (Thinking Machines Lab, Inc. (named on its model card page and in its training-content summary)): No small build.
- not this job: Inkling-Small (Thinking Machines Lab, Inc. (named on its model card page and in its training-content summary)): Smallest build 75 GB.
- not this job: InternVL3.5-GPT-OSS-20B-A4B-Preview (Shanghai AI Laboratory (OpenGVLab), not named in the repository): No publisher GGUF, no Ollama entry, no tool template and no thinking switch.
- not this job: Kimi K3 (Moonshot AI (the licence names Moonshot AI; the Kimi Platform terms name Moonshot AI PTE. LTD.)): No small build.
- not this job: Kolibri-1 (Aleph Alpha GmbH): Smallest third-party build 17.6 GB at IQ1_S and 49.8 GB at Q4; 78B total is a main model.
- not this job: Kronos (No company. Yu Shi publishes it from the GitHub account shiyu-coder and the Hugging Face user account NeoQuasar (named ShiYu), with co-authors at Tsinghua University per the paper): Not a language model; cannot answer routing or summary prompts.
- not this job: Laguna M.1 (Poolside, Inc. (named in Poolside's Terms of Use; the LICENSE.md copyright line reads Poolside)): 225B total parameters, 139 GB at the smallest; cannot sit beside a main model in spare memory.
- not this job: Laguna S 2.1 (Poolside, Inc. (named in Poolside's Terms of Use)): 96 GB at Q4_K_M; cannot be the second model in spare memory.
- not this job: LBS-27B (inclusionAI, Ant Group (named in the site footer of inclusion-ai.org); the work is credited to authors from Wuhan University of Technology, CASIA, Zhejiang University and Ant Group): BF16 only, 54.7 GB; not a small cousin and not a chat or routing model.
- not this job: MiniMax-M2.5 (MiniMax AI (named in the licences; MiniMax's platform terms name Nanonoble Pte. Ltd., Singapore)): Smallest build is 56 GB; cannot share memory with a main model.
- not this job: MiniMax-M2.7 (MiniMax (named in the licence's copyright line; MiniMax's platform terms name Nanonoble Pte. Ltd., Singapore)): Smallest build is 61 GB (UD-IQ1_M) and thinking cannot be turned off; it cannot sit beside a main model.
- not this job: MiniMax-M3.1-Flash (MiniMax (named in the licence's copyright line; MiniMax's platform terms name Nanonoble Pte. Ltd., Singapore)): A 428B model with a 129 GB smallest build; it cannot be the small cousin.
- not this job: Mistral Large 4 (Mistral AI (RCS Paris 952 418 325)): A 1T-parameter model; no build is small enough to sit beside a main model.
- not this job: Nemotron 3 Super 120B A12B (NVIDIA Corporation): Needs about 64 GB to 72 GB at 4 bits; it cannot be the small cousin.
- not this job: Nemotron 3 Ultra 550B A55B (NVIDIA Corporation): A 550B model; nothing small enough to sit beside a main model.
- not this job: Olmo 3 7B Think (The Allen Institute for Artificial Intelligence (Ai2), a 501(c)(3) nonprofit, per its Responsible Use Guidelines): Every reply opens a think block with no switch, so a routing or title call cannot skip reasoning.
- not this job: Ornith-1.5-397B (Ornith): 397B model.
- not this job: Qwen3.5-122B-A10B (Alibaba Cloud (named in the licence's copyright line)): At 34 GB even at IQ1_M it is a main model, not a cousin in 2 to 25 GB of spare memory.
- not this job: Qwen3.5-397B-A17B (Alibaba Cloud (named in the licence's copyright line)): Over 100 GB at any precision; not a cousin.
- not this job: Qwen3.6-27B (Alibaba Cloud (named in the licence's copyright line)): A 17 GB Q4 main model; acc-26 measured Qwen 3.6 at 26.7 s on routing prompts, far past 3 seconds.
- not this job: Qwen3.8-27B (Alibaba Cloud (named in the licence's copyright line)): A 17 GB Q4 main model with default xhigh reasoning; acc-39 pairs it with a smaller fast model rather than using it as the cousin.
- not this job: Qwen3.8-Flash-Next (Alibaba Cloud (Singapore) Private Limited, the registrant of qwen.ai. The licence names Qwen as copyright holder): Over 66 GB at any precision; not a cousin.
- not this job: Qwen3-Coder-Next (Alibaba Cloud (the Qwen team's employer; no LICENSE file names a copyright holder for this model)): 48 GB at Q4_K_M; a main model, not a cousin.
- not this job: Qwen3-Omni-30B-A3B-Instruct (Alibaba Cloud (carried from the 20 October 2026 pilot; this model's files name no copyright holder)): 18.6 GB at Q4 even in ggml-org's GGUF; not a cousin.
- not this job: Qwen3-VL-32B-Instruct (Alibaba Cloud (per the first batch's RDAP check of qwen.ai; this model's files name no copyright holder)): 19.8 GB at Q4 with 256 KB of KV per token; a main model.
- not this job: Qwopus3.8-27B-Flash-V2 (Jackrong, an individual Hugging Face account): A 17 GB Q4 model the card recommends with thinking on at xhigh effort; not a cousin.
- not this job: Recursive Flint (Recursive Superintelligence, Inc. (the Hugging Face organisation's full name)): Reasoning always on and 23 GB at INT4; a routing call spends thinking tokens first.
- not this job: Ripple-65B Preview (Ricursive Intelligence, Inc.): 57 GB NVFP4 and always-on thinking; not a cousin.
- not this job: Swift 1.5 Qwen3.8-27B (UkisAI): A 17 GB Q4 model built to think; not a cousin.
- not this job: Swift 1.5 Qwen3.8-Flash-Next (UkisAI): Over 66 GB at any precision; not a cousin.
- not this job: ThinkingCap (Qwen3.8 27B) (BottleCap AI (named in the licence's copyright line)): A 17 GB Q4 model tuned for shorter but still-on reasoning; not a cousin.
- not this job: Baidu Unlimited-OCR-2 (Baidu, Inc.): Not a chat model; it has no chat template and answers only page-parsing prompts.

28 listings in the category were added after this reading and aren't marked yet.

## What stops agents finishing

- thinking cannot be turned off, so a routing call takes 20 seconds
- no small quantised build, so the cousin takes the main model's memory

## Around it

Before: open-weights.local-coding-agent. After: coding-agents.plan-review.

## Where this comes from

Weight 8 of 100 in Open-weight models: 6 of 51 accounts pair a small model with a larger one for utility or planner roles

As data: https://www.anchorterminal.com/api/v1/use-cases/open-weights.sub-task-cousin.json, or the MCP tool get_use_case with id open-weights.sub-task-cousin. Draft, read 2026-11-12 from public docs and first-hand accounts, not yet reviewed by a person.

---

Query this catalogue directly over MCP: https://www.anchorterminal.com/mcp (Streamable HTTP, no key). How: https://www.anchorterminal.com/agents/#mcp
