Thinking Machines Lab Releases Inkling 2 Mini, Claims Top Open-Weight Coding Performance Under 100B Parameters
Mira Murati's startup Thinking Machines Lab released Inkling 2 Mini on November 20, 2026, an open-weights multimodal model it says leads every open-weights model under 100 billion parameters on coding benchmarks [7][2][5]. The 26.4-billion-parameter mixture-of-experts model uses a looped transformer architecture—its middle 12 layers run three times per token—to achieve 48 layer applications from 24 stored layers. The company reports 48.2% on DeepSWE v1.1 and 79.2% on SWE-bench Verified, ahead of Qwen3.8-27B's 42.2% and 77.8% respectively [1]. The weights are available under Apache 2.0 on Hugging Face and run locally on hardware with 32 GB or more of unified memory or VRAM [2][3][5][7].
Looped architecture trades memory for serial compute
Inkling 2 Mini's 24 decoder layers are divided into four blocks of six. The first and last blocks run once; the two middle blocks run three times with shared weights. Each token therefore sees 48 layer visits while the checkpoint stores only 24 layers. The model routes each token to 8 of 128 experts per layer, yielding 3.4 billion active parameters per pass and roughly 5.4 billion effective weight activations when the three loop passes are counted [1][3]. Most weights were trained and deployed natively in NVFP4 (4-bit) with the final four layers in MXFP8 (8-bit), avoiding post-training quantization loss [1]. The checkpoint is approximately 20 GB [1][8].
The company says looping increases serial reasoning depth without requiring the model to write out its chain of thought. In its own tests, a non-looped version cost 7.5% less compute to train but needed 140% more inference compute to match the looped model's loss on 512k-token agentic traces [1].
Benchmark claims are vendor-reported and harness-dependent
Thinking Machines' model card compares Inkling 2 Mini against Gemma 4 31B, Muse Glimmer, Qwen3.8-27B, and Nemotron 3.5 Super VL across 48 benchmarks [1][8]. Inkling leads or ties on most coding and agentic rows: DeepSWE v1.1 (48.2% vs 42.2%), SWE-bench Pro v2 Hard (77.9% vs 73.8%), SWE-bench Multilingual (73.7% vs 70.0%), SWE-bench Verified (79.2% vs 77.8%), Vibe Code Bench (77.9% vs 64.9%), and PMPP-Hard kernel optimization (455.82 vs 262.50) [1]. It ties Qwen on Terminal Bench v4.0 (5.6% each) and trails on Terminal Bench v2.1 (71.6% vs 73.0%), AutomationBench-AA (41.8% vs 48.2%), and EnterpriseOps-Gym (42.0% vs 44.2%) [1][8].
On the Artificial Analysis Intelligence Index v5.0, both models score 20.1 [1][4]. Independent testing by Artificial Analysis found Inkling about six times cheaper and four times faster per task than Qwen3.8-27B, though it uses roughly 38% more output tokens for the same score [7].
All coding scores come from Thinking Machines' own evaluation harnesses (mini-swe-agent v2.4.6, stirrup v0.2.1, pi v0.83.0, openhands). Harness choice can shift results materially; Qwen's own model card reports 61.7% on SWE-bench Pro under a different setup [5]. The SWE-bench Pro Hard subset contains only 51 tasks, so a 4.1-point gap represents roughly two additional solved tasks [5].
Hands-on testing reveals a gap on open-ended agentic debugging
Mindstudio.ai ran Inkling 2 Mini through a Hermes agent harness against a planted bug in a solar-panel installer app: a tiered electricity pricing calculation that charged all usage at the highest bracket. The model ran for over an hour without pinpointing the root cause, appearing to loop without converging. The same reviewer noted Qwen3.8 solved similar bugs in five to ten minutes [2]. The test was stopped early due to time constraints.
The model card itself acknowledges repetitive reasoning: "In internal testing the model sometimes became stuck in long reasoning loops, or repeatedly re-checked its own work without making progress, particularly when debugging code" [8]. It recommends setting output-token and agent-turn limits.
Factual recall and instruction following lag behind Qwen
Inkling 2 Mini's non-hallucination rate on AA-Omniscience is 63.4% versus Qwen3.8-27B's 71.4% [1][4]. The company explains that looping improves reasoning over supplied context but does not increase factual storage, which scales with total parameter count [1].
On ComplexConstraints v1.1 (many-constraint instructions), the model card's evaluation table shows Inkling at 33.4% and Qwen3.8-27B at 30.1%—Inkling ahead [1][8]. However, the same model card's "Limitations" section states the model scores 33.4% on ComplexConstraints v1.1 and 50.0% on MultiChallenge (multi-turn chat), "behind Qwen3.8-27B on both" [1]. The table does not show a Qwen score for MultiChallenge, so the comparison cannot be verified from the published data for that benchmark. Vision benchmarks are competitive: MMMU Pro 74.9% vs 76.3%, OmniDocBench 92.8% vs 91.5% [1][8].
Local deployment is practical but runtime support is new
The model runs on Apple silicon with 32 GB+ unified memory, NVIDIA DGX Spark, and GPUs with 32 GB+ VRAM [1][6][8]. GGUF quantizations (4-bit UD-Q4_K_XL at 17.6 GB, 8-bit Q8_0) are available via Unsloth, llama.cpp, Ollama, and LM Studio [1][3][6][7]. SGLang, vLLM, and MLX also support the architecture [9]. Context window is advertised at 1 million tokens, but each loop pass retains its own KV state, so practical context on consumer hardware will be lower [1][5].
API access is currently limited to Thinking Machines' Tinker platform and a single OpenRouter provider (Prime Intellect) at 73 tokens/second, compared to 24 providers for Qwen3.8-27B with fastest at 303–357 tokens/second [4].
Monitorability research released alongside
Because looping lets models reason internally without emitting tokens, Thinking Machines also released a variable-loop research variant (TML-d24-varloop-research) trained on 1–5 passes to study monitorability [1]. In preliminary tests, the no-CoT time horizon rose from 16 seconds at one pass to 37 seconds at three passes on the Think Fast benchmark. At three passes, chain-of-thought length fell from 3,703 to 1,731 tokens; at five passes it reached 1,661 tokens. In both cases the reasoning remained legible—sentence length and dictionary-word fraction were unchanged [1]. The model showed no improved ability to hide sabotage from a monitor at this scale [1]. The company calls results preliminary and notes the research model received RL only at three passes; deeper RL might compress reasoning further [1].
Bottom line
Inkling 2 Mini is a genuine architectural experiment: a looped MoE model that achieves competitive coding benchmarks at a size that fits on a high-end laptop. Its vendor-reported numbers are strong on agentic coding and search benchmarks, and independent cost/latency measurements are favorable. But the same looping that helps on serial reasoning also enables silent computation that monitors cannot see, and at least one hands-on agentic debugging test exposed a failure mode where the model looped for an hour without solving a realistic bug. Factual recall, multi-turn chat, and complex instruction following remain weaker than Qwen3.8-27B, though the model card's own table shows Inkling ahead on ComplexConstraints v1.1 while its narrative claims the opposite—a discrepancy readers should weigh. For developers choosing a local coding model today, Inkling 2 Mini is a credible new default for repository-level patching and tool-use workflows—but it is not a universal replacement, and its benchmark leads should be weighed against harness dependence, small test subsets, and the gap between controlled evaluations and open-ended agent loops.
Sources
- Introducing Inkling 2 Mini (thinkingmachines.ai)
- Inkling 2 Mini Benchmarks: How Thinking Machines' Open Model ... (mindstudio.ai)
- Inkling 2 Mini - Run Locally Guide (unsloth.ai)
- Inkling 2 Mini: Thinking Machines' 20GB Open Model Ties Qwen3.8 ... (neurohive.io)
- Inkling Mini 2 dethrones Qwen 3.8–27B for local coding | by Mehul Gupta (medium.com)
- Inkling 2 Mini (lmstudio.ai)
- Thinking Machines releases Inkling 2 Mini, a looped AI model for laptops (thenextweb.com)
- thinkingmachines/Inkling-2-Mini · Hugging Face (huggingface.co)
- Welcome Inkling 2 Mini, Cutting-edge Performance for your Local Setup (huggingface.co)