Key takeaways
- Inkling 2 Mini and Qwen3.8-27B both score 20 on the Artificial Analysis Intelligence Index, Qwen at its default xhigh effort.
- Inkling 2 Mini uses 5.4B active parameters counting its loops and ships as a 20GB checkpoint; the dense Qwen3.8-27B uses all 27B and its BF16 weights are about 56GB.
- Qwen3.8-27B costs $0.50 in and $3.00 out per million tokens on Qwen Cloud, while Inkling 2 Mini is listed at $0.15 and $0.50 on Tinker, as of November 20, 2026.
Inkling 2 Mini vs Qwen3.8: the two open-weight models score the same on the Artificial Analysis Intelligence Index, 20 each, but Inkling 2 Mini gets there with far less work per token. It uses 5.4 billion active parameters, counting its loops, against 27 billion for the dense Qwen3.8-27B, and its weights take about 20GB instead of about 56GB. Qwen3.8-27B is the one you can call today from two dozen API providers, has been out since August, and leads on office-style agent tasks and on avoiding wrong answers.
Thinking Machines Lab released Inkling 2 Mini on November 20, 2026. Alibaba's Qwen team released Qwen3.8-27B in August 2026. Both are Apache 2.0 and both read images. This comparison is based on the two models' docs, model cards and published benchmark results, read on November 20, 2026. We did not run the models ourselves.
Inkling 2 Mini vs Qwen3.8-27B at a glance
| Inkling 2 Mini | Qwen3.8-27B | |
|---|---|---|
| Maker | Thinking Machines Lab | Qwen (Alibaba) |
| Released | November 20, 2026 | August 2026 |
| Design | Mixture-of-experts, looped | Dense |
| Total parameters | 26.4B | 27B |
| Active per token | 3.4B (5.4B counting loops) | 27B |
| Weights on Hugging Face | About 20GB (NVFP4 and MXFP8) | About 56GB (BF16) |
| Context window | Up to 1M tokens | 262,144 native, up to 1M with YaRN |
| Inputs | Text, images | Text, images, video |
| License | Apache 2.0 | Apache 2.0 |
| AA Intelligence Index | 20 | 20 (xhigh effort) |
| AA output speed | 280 tokens/s | 53 tokens/s (xhigh) |
| Hosted price per 1M tokens | $0.15 in, $0.50 out (Tinker sampling) | $0.50 in, $3.00 out (Qwen Cloud) |
How the two models are built
The difference in design explains most of the rest of this comparison.
Qwen3.8-27B is a dense model: all 27 billion parameters work on every token. Its model card lists 64 layers in 16 blocks, each with three Gated DeltaNet layers (a linear-attention layer that keeps a fixed-size running memory instead of looking back at every earlier token) and one regular attention layer. Thinking is on by default, and a reasoning_effort setting picks xhigh, medium or low.
Inkling 2 Mini is a mixture-of-experts model, which means each token goes to only a few of many small sub-networks: 8 of 128 experts per layer. It stores 24 layers, and the middle 12 run three times for each token. That gives 48 layer passes from 24 layers of weights. Thinking Machines puts the active count at 3.4 billion parameters, or 5.4 billion once the repeated passes are counted. Our news story on the launch explains the looping in more detail.
The trade is simple. Inkling 2 Mini spends compute by reusing layers instead of storing more of them. That keeps the download small and each token cheap. The cost is knowledge: the Inkling 2 Mini model card says loops do not help on knowledge-heavy tests, because what a model knows depends on its total size.
Memory and hardware you need
Inkling 2 Mini ships as a single checkpoint of about 20GB. Thinking Machines trained it in two low-precision number formats, NVFP4 for 20 layers and MXFP8 for the last four, so there is no separate quantization step that loses quality. The model card lists one NVIDIA DGX Spark, or an Apple silicon Mac with 32GB or more of unified memory, and recommends 32GB or more of VRAM on other GPUs.
Qwen3.8-27B's official weights are BF16, and the safetensors files on Hugging Face add up to about 56GB. The model card does not give a memory requirement. It points to quantized builds for llama.cpp, Ollama and LM Studio for smaller machines, and to vLLM, SGLang or TokenSpeed for production serving.
Speed follows from the active parameter count. Artificial Analysis measures Inkling 2 Mini at 280 output tokens per second and Qwen3.8-27B at 53 tokens per second at its default xhigh effort. Those are hosted API speeds, not laptop speeds. Thinking Machines adds that support for looped models is new in local runtimes, so speed varies by runtime and hardware.
Benchmarks: a split decision
The independent number is the Artificial Analysis Intelligence Index, which combines many tests run the same way on every model. Both models score 20. Qwen3.8-27B gets its 20 at xhigh effort. At medium effort it drops to 17, at low effort to 16, and with reasoning off to 10.
The detailed scores below come from Thinking Machines' own table, which ran both models in the same harnesses. Treat them as the lab's numbers. Two of the Qwen scores, 73.0% on Terminal Bench 2.1 and 42.2% on DeepSWE 1.1, match what Qwen reports on its own card.
| Benchmark | Inkling 2 Mini | Qwen3.8-27B |
|---|---|---|
| SWE-bench Verified (bash only) | 79.2% | 77.8% |
| SWE-bench Pro v2 (Hard) | 77.9% | 73.8% |
| SWE-bench Multilingual | 73.7% | 70.0% |
| DeepSWE v1.1 | 48.2% | 42.2% |
| Terminal Bench v2.1 | 71.6% | 73.0% |
| Vibe Code Bench v1.1 | 77.9% | 64.9% |
| BrowseComp | 79.0% | 61.6% |
| MCP Atlas | 82.7% | 77.3% |
| GDPVal-AA v2.1 (Elo) | 1319 | 1427 |
| AutomationBench-AA v1.0.6 | 41.8% | 48.2% |
| MMMU Pro | 74.9% | 76.3% |
| HLE-Diamond | 15.2% | 15.3% |
| LongBench v2 | 66.9% | 52.3% |
| AA-Omniscience non-hallucination rate | 63.4% | 71.4% |
Inkling 2 Mini leads on coding agents that fix real repositories, on web research (BrowseComp) and tool use (MCP Atlas), and on long documents. Qwen3.8-27B leads on office work such as GDPVal and AutomationBench, on the image-reasoning test MMMU Pro, and on the AA-Omniscience non-hallucination rate, which means fewer confident wrong answers. Inkling 2 Mini leads on most of the other image tests in the table, including OmniDocBench and ScreenSpot Pro.
Thinking Machines also reports internal safety tests in which Inkling 2 Mini took destructive actions in 10.2% of cases against 35.2% for Qwen3.8-27B. Its own model card still warns that the model may delete or overwrite files without asking and can get stuck in repetitive reasoning loops while debugging, so sandbox it for agent work.
Qwen's card does not include Inkling 2 Mini, which came out three months later. Its own headline numbers include 89.2% on GPQA Diamond and 84.3% on OSWorld-Verified, tests that the Inkling table does not report.
Where to run each model and what it costs
Qwen3.8-27B is easy to reach. Qwen Cloud charges $0.50 per million input tokens and $3.00 per million output tokens, with a 1M-token context window, as of November 20, 2026. OpenRouter lists 24 providers for it, with input prices from $0.035 to $0.99 per million tokens and output from $1.05 to $3.20. Most of those providers cap context at 262,144 tokens.
Inkling 2 Mini is harder to buy as an API on launch day. It is not on OpenRouter. Thinking Machines serves it on Tinker, its platform for inference and fine-tuning, and the Tinker pricing page lists it at $0.15 per million prefill (input) tokens and $0.50 per million sampled (output) tokens, with training at $0.75 per million tokens at 64K context. That page also says pay-as-you-go serverless inference covers only Inkling 2 Small, Inkling and Inkling-Small for now, with a waitlist for production use. Artificial Analysis lists the same $0.15 and $0.50 prices.
For local use, both models work with Transformers, vLLM, SGLang, llama.cpp, Ollama, LM Studio and Docker Model Runner. Unsloth has GGUF builds of Inkling 2 Mini.
Which one to pick
Pick Inkling 2 Mini if you want to run a capable coding or research agent on one machine with 32GB of memory, or if you pay per token and output speed matters. Its benchmark lead on repository-level coding is real in the lab's table, and the Artificial Analysis tie shows it is not a weaker model overall.
Pick Qwen3.8-27B if you need a hosted API from several providers today, video input, stronger results on office tasks, fewer confident wrong answers, or good output in Chinese. Inkling 2 Mini's card says its quality drops outside English, with mixed-language and clipped phrasing seen in Chinese.
The open question is how Inkling 2 Mini holds up outside Thinking Machines' harnesses. It is one day old, and so far the only independent number is the Artificial Analysis index. Watch for its results on outside coding leaderboards. If you like the Inkling family but need more factual recall, the model card points to the larger Inkling 2 Small as the stronger model on knowledge.
Frequently asked questions
Is Inkling 2 Mini better than Qwen3.8-27B?
Overall they tie: both score 20 on the Artificial Analysis Intelligence Index. In Thinking Machines' own tests Inkling 2 Mini leads on coding agents, web research and long documents, while Qwen3.8-27B leads on office tasks, MMMU Pro and avoiding wrong answers.
How much memory does Inkling 2 Mini need?
The checkpoint is about 20GB. Thinking Machines lists an Apple silicon Mac with at least 32GB of unified memory, one NVIDIA DGX Spark, or a GPU with 32GB or more of VRAM.
Can Qwen3.8-27B run on one GPU?
Its official BF16 weights are about 56GB, so the full-precision model needs more than one consumer GPU. Qwen's model card points to quantized builds for llama.cpp, Ollama and LM Studio for smaller machines, without giving a memory figure.
Are Inkling 2 Mini and Qwen3.8-27B free to use commercially?
Both are released under the Apache 2.0 license, which allows commercial use. Inkling 2 Mini also comes with a Thinking Machines acceptable use policy.
Sources
- Inkling 2 MiniThinking Machines Lab
- thinkingmachines/Inkling-2-Mini model cardThinking Machines Lab
- Qwen/Qwen3.8-27B model cardQwen
- Qwen3.8-27B on Qwen CloudQwen Cloud
- Tinker pricingThinking Machines Lab
- Inkling 2 Mini: intelligence, performance and priceArtificial Analysis
- Qwen3.8 27B: intelligence, performance and priceArtificial Analysis
- Qwen3.8 27B providers and pricesOpenRouter

