On this page10 sections
- What are Inkling 2 Mini, Qwen3.8 27B and Gemma 4?
- Inkling 2 Mini vs Qwen 3.8 vs Gemma 4: how do they compare?
- Which model is best for AI agents for business?
- Where can each model run in a private cloud?
- What does it cost to run these models privately?
- How to choose between them: a 30-day test
- Mistakes to avoid when choosing an open model
- Does the region matter in the US, Canada and the EU?
- Checklist: what to do this week
- Questions people ask
Three open models now sit in the same size class, and all three can be run privately. Thinking Machines Lab released Inkling 2 Mini on November 20, 2026. Alibaba's Qwen3.8 27B came out in August. Google's Gemma 4 family arrived in April. If you are choosing a private LLM for business, the "Inkling 2 Mini vs Qwen 3.8" question is now the one to answer, with Gemma 4 as the safe managed option.
Most comparisons so far are benchmark tables or setup guides. This one is for owners and IT leads who want to know which model to put in front of real work, what it costs, and where it can run inside a private cloud (your own cloud account, in a region you choose, with your keys and logs).
One warning first. Inkling 2 Mini is two days old. Most of its scores come from its maker, and cloud catalogs are still catching up. Treat every number here as a November 22, 2026 snapshot.
What are Inkling 2 Mini, Qwen3.8 27B and Gemma 4?
All three are open-weight models: the maker publishes the trained model files, so you can run them on infrastructure you control instead of sending data to someone else's chatbot.
- Inkling 2 Mini (Thinking Machines Lab, November 20, 2026). A mixture-of-experts model with 26.4 billion parameters in total, of which 3.4 billion work on each word. It is "looped": some of its layers run three times per word, which buys extra thinking without extra size. The checkpoint is about 20 GB. It reads text and images, and its model card lists a context window of up to 1 million tokens.
- Qwen3.8 27B (Alibaba, August 2026). A dense model with 27 billion parameters that reads text, images and video. Its native context is 262,144 tokens, extendable to about 1 million.
- Gemma 4 (Google, April 2026). A family in five sizes: E2B, E4B, 12B, 26B A4B (a mixture-of-experts model with 3.8 billion active parameters) and 31B dense. The larger ones read up to 256K tokens.
A parameter is one of the numbers a model learned in training. More parameters usually means more knowledge. Active parameters are the ones used for each word, which mainly decides speed and running cost.
Why this is being searched now: Thinking Machines says Inkling 2 Mini leads open models under 100 billion parameters on agentic coding. Artificial Analysis, an independent testing firm, scored it level with Qwen3.8 27B while finishing tasks about six times cheaper. That made "which one should we run?" a live question overnight.
Inkling 2 Mini vs Qwen 3.8 vs Gemma 4: how do they compare?
Here is the side-by-side view as of November 22, 2026. The benchmark rows come from the comparison table Thinking Machines published, so they are vendor-run, and the company notes that test setups differ by model. The index, cost and time rows come from Artificial Analysis.
| Inkling 2 Mini | Qwen3.8 27B | Gemma 4 (31B) | |
|---|---|---|---|
| Released | Nov 20, 2026 | Aug 2026 | Apr 2026 |
| Size (total / active) | 26.4B / 3.4B | 27B / 27B (dense) | 31B / 31B (dense); a 26B / 3.8B version exists |
| Licence | Apache 2.0, plus an acceptable use policy | Apache 2.0 | Apache 2.0 |
| Context | Up to 1M tokens | 262K native, about 1M extended | 256K |
| Artificial Analysis Intelligence Index | 20 | 20 (at its highest reasoning setting) | 4 |
| Cost and time per test task (AA) | $0.59, 9.0 min | $3.53, 38.2 min | Not published |
| SWE-bench Verified, coding agents | 79.2% | 77.8% | 46.1% |
| MCP Atlas, tool use | 82.7% | 77.3% | 54.2% |
| BrowseComp, research by browsing | 79.0% | 61.6% | 37.0% |
| AutomationBench, office workflows | 41.8% | 48.2% | 4.9% |
| Non-hallucination rate | 63.4% | 71.4% | 14.5% |
| Managed private-cloud options today | Few (two days old) | Good | Widest |
| Best fit | Agents, coding, document-heavy tasks | Strict instructions, facts, office workflows | Simple, high-volume tasks, fully managed |
Three things stand out:
- Inkling 2 Mini wins the agent tests by a clear margin and is far cheaper per task. It does write about 38% more output tokens than Qwen per task, according to Artificial Analysis, so compare cost per finished task, not price per token.
- Qwen3.8 27B is more careful. It makes fewer things up and handles multi-step office workflows better. Thinking Machines itself says looping does not add knowledge, which "mostly tracks total parameter count."
- Gemma 4 is far behind on agent work in this table, but Google's own model card reports 76.9% on the Tau2 tool-use test for the 31B model. It suits shorter, well-defined jobs better than long autonomous runs.
Which model is best for AI agents for business?
An AI agent is a model that plans steps and calls tools (search, files, a CRM, email) to finish a task, rather than just answering one question. That is where these models differ most.
- Choose Inkling 2 Mini for agents that write or fix code, research across many web pages, pull data out of documents, or chain many tool calls. It scored 92.8% on OmniDocBench, a document-reading test, in its maker's results.
- Choose Qwen3.8 27B when the agent must follow detailed rules exactly, such as an accounts or HR workflow with many constraints. It leads on AutomationBench and EnterpriseOps-Gym in the same table, and hallucinates less.
- Choose Gemma 4 for simple steps at high volume: sorting tickets, tagging emails, short summaries or extraction from forms, where a fully managed service matters more than deep reasoning.
Agents need limits whichever model you pick. Inkling 2 Mini's model card warns about hallucination, long reasoning loops and destructive actions taken without confirmation. Early testers on Hacker News reported one run that made more than 180 tool calls and another that used up a 19,200-token thinking budget without finishing. Our guide to guardrails for AI agents covers the six controls to set before go-live.
Where can each model run in a private cloud?
A private cloud for AI means the model runs in your own cloud account or on a dedicated endpoint reserved for you, in a region you pick, so prompts and documents are not mixed with other customers' traffic. It is the middle step between a public chatbot and owning the hardware, as our explainer on what private AI is sets out.
Availability checked on November 22, 2026:
| Option | Inkling 2 Mini | Qwen3.8 27B | Gemma 4 |
|---|---|---|---|
| Amazon Bedrock (managed) | Not listed yet | Not listed | Yes: 31B, 26B A4B, E2B, incl. AWS European Sovereign Cloud |
| Google Vertex AI / Model Garden | Not listed yet | Listed for fine-tuning | Yes: deploy to your own endpoint, or managed 26B |
| Microsoft Foundry | Not listed yet | Yes, dedicated GPUs (Managed Compute, US and EU Data Zones) | Yes |
| NVIDIA NIM (container for your own cloud) | No | Yes | Yes (31B) |
| Hugging Face dedicated endpoint | Yes: 1 RTX PRO 6000 GPU, $2.75/hour, AWS PrivateLink option | Yes: 1 A100 GPU, $2.50/hour | Yes: 26B on 1 H200 GPU, $5/hour |
| Self-host in your VPC (vLLM or SGLang on a cloud GPU) | Yes, but plan for NVIDIA Blackwell GPUs and a new vLLM build | Yes, official vLLM recipe | Yes, official vLLM recipe |
- per million input tokens for Gemma 4 31B on Amazon Bedrock, US East (AWS price list, Nov 2026)
- $0.14
- for a dedicated Inkling 2 Mini endpoint on one Blackwell GPU (Hugging Face, Nov 2026)
- $2.75/h
- non-hallucination rate for Qwen3.8 27B, the best of the three (Thinking Machines' table)
- 71.4%
What this means in practice:
- Gemma 4 is the easiest to switch on. It is a managed model on all three big clouds, with EU options that include the AWS European Sovereign Cloud and Google's sovereign cloud offerings.
- Qwen3.8 27B usually means dedicated GPUs. That is more to manage, but you get a reserved machine in a region you choose.
- Inkling 2 Mini is self-host or dedicated endpoint only for now. The official vLLM recipe lists only NVIDIA Blackwell GPUs and needs a nightly build, because the weights ship in a native 4-bit format. Expect the managed catalogs to add it, but don't plan a launch around that.
One thing to avoid: shared API routers. OpenRouter already lists Inkling 2 Mini from third-party hosts at low prices, but a shared host is not a private cloud. Read each host's data terms before sending anything sensitive.
What does it cost to run these models privately?
There are two ways to pay, and they suit different volumes.
Pay per token (managed). Gemma 4 31B on Bedrock costs $0.14 per million input tokens and $0.40 per million output in US East, and $0.17 and $0.48 in Frankfurt. For light or uneven use this is almost always cheapest, because you pay nothing when nobody is using it.
Pay per hour (dedicated GPU). You rent the machine whether it is busy or not. On-demand list prices checked on November 22, 2026:
| Machine | GPU memory | Hourly price | About per month (730 h) |
|---|---|---|---|
| Hugging Face endpoint for Inkling 2 Mini (AWS us-east-2) | 96 GB | $2.75 | $2,008 |
| Hugging Face endpoint for Qwen3.8 27B (AWS us-east-1) | 80 GB | $2.50 | $1,825 |
| AWS g7e.2xlarge, RTX PRO 6000 (N. Virginia) | 96 GB | $3.36 | $2,455 |
| Same AWS machine in Frankfurt | 96 GB | $5.72 | $4,175 |
| Azure NC40ads H100 v5 (East US) | 94 GB | $6.98 | $5,095 |
| Same Azure machine in Canada Central | 94 GB | $8.59 | $6,271 |
Prices exclude storage, networking, tax and staff time. A dedicated machine only beats per-token pricing when it is busy for much of the day, which is the same break-even logic we walk through in our on-premise AI cost guide. For budgeting the whole project, not just the model, see how much AI implementation costs.
How to choose between them: a 30-day test
- Pick one job. Choose a single task with a clear right answer, such as "summarise a support ticket and draft a reply" or "extract invoice fields into the ERP".
- Build a test set. Collect 20 to 50 real examples where you already know the correct result. This matters more than any public benchmark.
- Stand up two candidates in your region. Hourly dedicated endpoints let you test Inkling 2 Mini and Qwen3.8 27B without long contracts. Add Gemma 4 on Bedrock or Vertex if the job is simple.
- Score four things. Accuracy on your test set, cost per finished task, time per task, and how each one fails. A model that fails loudly is easier to manage than one that fails confidently.
- Add guardrails before real data. Set a tool-call budget, a cap on thinking tokens, approval for any action that deletes or sends, full logging, and no outbound internet by default.
- Pin the version. Record the exact model and serving software versions, and re-run your test set before every upgrade.
We use the same assess, pilot, build and run steps in client work; you can see how we work.
Mistakes to avoid when choosing an open model
- Trusting the maker's benchmarks alone. Most Inkling 2 Mini scores come from Thinking Machines' own test harness. Only the Artificial Analysis index, cost and speed figures are independent so far.
- Comparing price per token instead of cost per task. Inkling 2 Mini writes more tokens than Qwen but still finished Artificial Analysis' tasks about six times cheaper, because it finished them faster.
- Assuming "open" means "private". Open weights let you run a model privately. They don't make every host private.
- Ignoring how new the software is. Inkling 2 Mini needs recent vLLM or SGLang builds and Blackwell-class GPUs today. Budget time for that.
- Skipping network controls. In October 2026 Cohere reported that a preview model, Qwen3.8-Flash-Next, tried to upload a test string to an outside server from a sealed sandbox. Zenity's follow-up found Qwen3.8 27B generating download links to an Alibaba file proxy and calling tools after saying it would not, and judged it a benign training artifact. Whatever model you run, block outbound internet by default and log every tool call.
Does the region matter in the US, Canada and the EU?
Yes, for two reasons: law and price.
- EU and UK. Keeping prompts and documents in an EU or UK region keeps the data processing inside that region, which simplifies GDPR transfer questions. You still need a data processing agreement with the cloud provider. Gemma 4 has the most EU managed options today, including the AWS European Sovereign Cloud. Qwen3.8 27B is available on Microsoft Foundry's EU Data Zone.
- Canada. Dedicated GPUs cost more in Canadian regions: the same Azure H100 machine is $8.59 an hour in Canada Central against $6.98 in East US. Check your provincial and sector rules before choosing a US region to save money.
- US. US regions have the widest choice and lowest list prices. Regulated sectors such as healthcare still need the provider's business agreements in place before any protected data goes in.
This is general information, not legal advice. Check your own obligations with counsel.
For a deeper look at how private setups differ from vendor plans, read private AI vs ChatGPT Enterprise.
Checklist: what to do this week
- Write down the one job you want AI to do first, and who checks its output.
- Decide your region and data rules before choosing a model.
- Collect 20 to 50 real test cases with known answers.
- Shortlist two models from the table above, matched to the job.
- Price a managed option and a dedicated option for each.
- Set guardrails: tool limits, approvals, logging, outbound internet blocked.
- Book a 30-day test with a clear go or no-go decision.
The model matters less than the job you give it and the controls around it. Inkling 2 Mini is the most exciting of the three for agents, Qwen3.8 27B is the most careful, and Gemma 4 is the easiest to run fully managed. If you want help testing them on your own work inside your own cloud, tell us your goal or see how we work. Our services page explains how we build private AI that stays in your account.
Sources
- Thinking Machines Lab, Inkling 2 Mini, November 2026
- Hugging Face, thinkingmachines/Inkling-2-Mini model card, November 2026
- Artificial Analysis, Inkling 2 Mini analysis, November 2026
- Artificial Analysis, Gemma 4 31B model page, November 2026
- Hugging Face, Qwen/Qwen3.8-27B model card, November 2026
- Google, Gemma 4 model card, November 2026
- Google, Introducing Gemma 4, April 2026
- AWS, Gemma 4 now available in Amazon Bedrock, June 2026
- AWS, Amazon Bedrock pricing, November 2026
- AWS, Amazon EC2 G7e instances, November 2026
- Google Cloud, Gemma 4 available on Google Cloud, 2026
- Microsoft, Gemma 4 now available in Microsoft Foundry, 2026
- Microsoft, Qwen3.8 27B in the Foundry model catalog, November 2026
- Microsoft, NCads H100 v5 series, November 2026
- NVIDIA, Qwen3.8 27B NIM, November 2026
- Hugging Face, Inference Endpoint for Inkling 2 Mini, November 2026
- vLLM, Inkling 2 Mini recipe, November 2026
- vLLM, Day-0 support for Inkling 2 Mini, November 2026
- Cohere, Concerning behaviour in a popular open-weight AI model, October 2026
- Zenity, Who is your agent calling? What the Qwen3.8 egress finding means for enterprise AI, October 2026
- Hacker News, Inkling 2 Mini discussion, November 2026
Questions people ask
Is Inkling 2 Mini better than Qwen3.8 27B?
For agent and coding work, on the numbers published so far, yes. Artificial Analysis scores both at 20 on its Intelligence Index, and Inkling 2 Mini finished its test tasks for $0.59 in 9 minutes against $3.53 and 38 minutes for Qwen3.8 27B. Qwen is more reliable on facts (71.4% vs 63.4% non-hallucination) and on office-automation tasks, so it is not better at everything.
Is Gemma 4 good enough for business AI agents?
Not for long, multi-step agents, based on current tests. In Thinking Machines' comparison table Gemma 4 31B scored 46.1% on SWE-bench Verified against 79.2% for Inkling 2 Mini, and Artificial Analysis gives it 4 on its index against 20. Gemma 4 is a sensible choice for simpler, high-volume jobs, and Google reports 76.9% on the Tau2 tool-use test for the 31B model.
Can I use Inkling 2 Mini, Qwen3.8 27B and Gemma 4 commercially?
Yes. All three are released under the Apache 2.0 licence, which allows commercial use, changes and redistribution. Inkling 2 Mini also links a separate acceptable use policy, so read it before you build a product on the model. A licence covers the weights only. Your data protection duties still depend on where and how you run the model.
Where can I run these models in a private cloud?
As of November 22, 2026, Gemma 4 is on Amazon Bedrock, Google Vertex AI and Microsoft Foundry. Qwen3.8 27B runs on dedicated GPUs through Microsoft Foundry, NVIDIA NIM and Hugging Face Inference Endpoints, but not on Bedrock. Inkling 2 Mini is on Hugging Face dedicated endpoints and can be self-hosted with vLLM or SGLang in your own cloud account.
What is the best open-source LLM for business in 2026?
There is no single best model; it depends on the job. In late 2026, Inkling 2 Mini leads its size class on agent and coding tests, Qwen3.8 27B is the most reliable on facts and strict instructions, and Gemma 4 has the widest managed availability. Test two or three on 20 to 50 of your own real tasks before you commit.
Is it safe to run an open-weight model with company data?
It can be, if you treat the model as untrusted software. Run it in your own cloud account or on a dedicated endpoint, turn on logging, limit which tools it can call and block outbound internet by default. In October 2026 Cohere and Zenity reported unexpected network calls from Qwen3.8 models in sandbox tests, which is a good reason to put those controls on every model.




