Innovation & Tech

Thinking Machines' Inkling 2 Mini Claims Coding Crown Under 100B

Mira Murati's Thinking Machines Lab ships a 26.4B looped MoE model claiming top sub-100B coding scores — but vendor harnesses, a tiny benchmark subset and one failed debugging test complicate the story.

By Olivia Hart

5 min read

Thinking Machines Lab Releases Inkling 2 Mini, Claims Top Open-Weight Coding Performance Under 100B Parameters
Thinking Machines Lab Releases Inkling 2 Mini, Claims Top Open-Weight Coding Performance Under 100B ParametersHarshit Katiyar / unsplash

What's News

  • Inkling 2 Mini released November 20, 2026: 26.4B-parameter looped MoE model, Apache 2.0, ~20 GB checkpoint
  • Reports 79.2% on SWE-bench Verified vs Qwen3.8-27B's 77.8%; all coding scores from Thinking Machines' own harnesses
  • Independent hands-on test: model looped over an hour on a planted pricing bug that Qwen3.8 solved in 5-10 minutes
  • Artificial Analysis found it ~6x cheaper and ~4x faster per task than Qwen3.8-27B, using ~38% more output tokens
  • Runs on 32 GB+ hardware; API access limited to two providers at 73 tokens/second vs 24 providers for Qwen

By Olivia Hart

Thinking Machines Lab released Inkling 2 Mini on November 20, 2026, a 26.4-billion-parameter open-weights model it says outperforms every open-weights model under 100 billion parameters on coding benchmarks. The claim rests almost entirely on the company's own harnesses, and at least one independent test showed the model looping for over an hour on a realistic bug. Mira Murati's startup is selling an architectural experiment; buyers should read the fine print.

The model 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. Weights are free under Apache 2.0 on Hugging Face and run locally on hardware with 32 GB or more of unified memory or VRAM. The checkpoint is roughly 20 GB.

What does the looped architecture actually do?

Inkling 2 Mini's 24 decoder layers split into four blocks of six. The first and last blocks run once; the two middle blocks run three times with shared weights, so each token 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 across the loop passes.

Most weights were trained natively in NVFP4 (4-bit), with the final four layers in MXFP8 (8-bit), which the company says avoids post-training quantization loss. In its own tests, a non-looped version cost 7.5% less to train but needed 140% more inference compute to match the looped model's loss on 512k-token agentic traces. The trade: deeper serial reasoning without writing out a chain of thought.

How solid are the benchmark claims?

The model card compares Inkling against Gemma 4 31B, Muse Glimmer, Qwen3.8-27B, and Nemotron 3.5 Super VL across 48 benchmarks. Inkling leads on most coding and agentic rows: SWE-bench Pro v2 Hard (77.9% vs 73.8%), SWE-bench Multilingual (73.7% vs 70.0%), Vibe Code Bench (77.9% vs 64.9%), and PMPP-Hard kernel optimization (455.82 vs 262.50). It ties Qwen on Terminal Bench v4.0 at 5.6% each and trails on AutomationBench-AA (41.8% vs 48.2%) and EnterpriseOps-Gym (42.0% vs 44.2%). On the Artificial Analysis Intelligence Index v5.0, both score 20.1.

Caveats abound. All coding scores come from Thinking Machines' own harnesses, and harness choice shifts results materially — Qwen's own card reports 61.7% on SWE-bench Pro under a different setup. The SWE-bench Pro Hard subset contains only 51 tasks, so a 4.1-point gap equals roughly two additional solved tasks. Independent testing by Artificial Analysis found Inkling about six times cheaper and four times faster per task than Qwen3.8-27B, though it burns roughly 38% more output tokens for the same score.

What happened in real-world testing?

Mindstudio.ai ran the model through a Hermes agent harness against a planted bug in a solar-panel installer app: a tiered electricity pricing calculation charging all usage at the highest bracket. The model ran for over an hour without pinpointing the root cause, appearing to loop without converging; Qwen3.8 solved similar bugs in five to ten minutes. The test was stopped early.

The model card admits the problem: "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." It recommends output-token and agent-turn limits.

Factual recall also lags: 63.4% non-hallucination on AA-Omniscience versus Qwen's 71.4%. The company explains looping improves reasoning over supplied context but does not increase factual storage, which scales with total parameter count. A discrepancy sits inside the card itself: its table shows Inkling ahead on ComplexConstraints v1.1 (33.4% vs 30.1%), while its Limitations section calls the model "behind Qwen3.8-27B on both" ComplexConstraints and MultiChallenge (50.0%). The table omits a Qwen MultiChallenge score, so that comparison cannot be verified. Vision results are competitive: MMMU Pro 74.9% vs 76.3%, OmniDocBench 92.8% vs 91.5%.

Is local deployment practical?

The model runs on Apple silicon with 32 GB+ unified memory, NVIDIA DGX Spark, and GPUs with 32 GB+ VRAM. GGUF quantizations — a 4-bit UD-Q4_K_XL at 17.6 GB and an 8-bit Q8_0 — are available via Unsloth, llama.cpp, Ollama, and LM Studio; SGLang, vLLM, and MLX also support the architecture. The advertised 1-million-token context carries a catch: each loop pass retains its own KV state, so practical context on consumer hardware will be lower. API access is thin — Thinking Machines' Tinker platform and a single OpenRouter provider at 73 tokens/second, versus 24 providers for Qwen3.8-27B topping out at 303–357 tokens/second.

Alongside the release, the company published a variable-loop research variant (TML-d24-varloop-research) trained on 1–5 passes to study monitorability. Preliminary results: the no-CoT time horizon rose from 16 seconds at one pass to 37 seconds at three passes on the Think Fast benchmark, and chain-of-thought length fell from 3,703 tokens to 1,731 at three passes while remaining legible. The model showed no improved ability to hide sabotage from a monitor at this scale; the company calls the results preliminary.

Inkling 2 Mini is a credible new default for local repository-level patching and tool use — but its benchmark leads depend on vendor harnesses, small test subsets, and controlled evaluations that one hour-long debugging failure already undercut.

Source: epyhia.com

Share this article:

More from Olivia Hart

Olivia Hart

Show full bio

Staff writer covering people, launches and day-to-day practice in entrepreneurship, startups and venture funding at Business Bearings.

2,535 articles

Related articles

« Previous article