Key takeaways

  • Thinking Machines Lab released Inkling 2 Mini on November 20, 2026, with open weights and Tinker inference and customization access.
  • The text-and-image model has 26.4B total parameters. Reusing its middle layers increases computation without storing another copy of those weights.
  • The model card describes a roughly 20 GB mixed-precision checkpoint and recommends Apple silicon with at least 32 GB memory or other GPUs with at least 32 GB VRAM.
  • OpenRouter lists thinkingmachines/inkling-2-mini through PrimeIntellect with a 1,048,576-token context and paid token rates. Listed access is not a completed inference test.
01

What launched and what is available now

Thinking Machines Lab released Inkling 2 Mini on November 20, 2026: an open-weight multimodal model designed for local computers, coding assistants and tool-using applications. The developer publishes thinkingmachines/Inkling-2-Mini on Hugging Face under Apache 2.0 and says it is available for inference and customization on Tinker, including chat through the Tinker Playground. It accepts text and images and generates text, not images or audio.

On November 21, OpenRouter added the managed route thinkingmachines/inkling-2-mini. Its registry creation timestamp is 02:19:45 UTC, distinct from the developer’s November 20 release date. At retrieval, the provider endpoint listed PrimeIntellect, FP4 serving, a 1,048,576-token context and a 65,536-token maximum completion. Tools, structured outputs and reasoning controls are advertised; we did not exercise them.

The listed price is $0.10 per million input tokens and $0.40 per million output tokens, with cached input reads at $0.02 per million tokens. This is a paid route, not a free-model test candidate. Those are OpenRouter’s observed rates, not a Tinker price quotation. Provider availability, account limits and rates can change; a listing is not proof of a successful authenticated inference call.

02

How a smaller checkpoint does more work

Inkling 2 Mini stores 24 decoder layers, but runs the middle 12 three times with shared weights. The result is 48 layer applications per token without storing 48 distinct layers. Its mixture-of-experts router selects eight of 128 experts at each layer. The card gives 26.4 billion total parameters and 3.4 billion active parameters, while counting repeated passes as 5.4 billion effective active parameters per token.

That distinction matters when buying hardware or comparing speed. Shared weights reduce the checkpoint footprint; they do not make the additional computation disappear. Hugging Face’s integration notes also say each decoder visit retains its own cache and short-convolution state. A million-token context specification is not evidence that a full million-token workload will fit comfortably on a minimum-memory laptop.

Thinking Machines distributes the checkpoint in mixed NVFP4 and MXFP8 precision at about 20 GB. Its card lists NVIDIA DGX Spark and Apple silicon with 32 GB or more memory, and recommends at least 32 GB VRAM for other GPUs. These are vendor requirements and recommendations, not memory measurements from AccessAllGPT. Allow additional room for caches, the runtime and other applications.

03

Local runtime support is part of the release

This is not a promise of weights arriving later. We verified an ungated public Hugging Face repository containing a safetensors model file, an MTP file, configuration, tokenizer and chat-template artifacts. We did not download those weights or validate their integrity. The repository metadata and model card declare Apache 2.0; the card separately links a model acceptable-use policy that deployers should review.

Hugging Face documents Transformers, llama.cpp/Llama App, SGLang and MLX integration. Its local CLI example uses llama serve -hf ggml-org/Inkling-2-Mini-GGUF:IQ4_NL. That is a publisher-documented command, not one we executed, and it selects a separate quantized repository rather than the native checkpoint. The SGLang example requires SM103 hardware; do not treat it as a universal GPU recipe.

The release also includes TML-d24-varloop-research, a distinct research variant with a configurable number of passes. Inkling 2 Mini itself uses fixed looping. Changing an API reasoning-effort setting is not the same as selecting fewer architectural passes, and results from the variable-loop artifact should not be silently attributed to the shipping model.

04

Promising coding results, with explicit reliability limits

Thinking Machines reports 48.2% on DeepSWE v1.1 and 77.9% on SWE-bench Pro v2 Hard. These are vendor-reported benchmark results, not AccessAllGPT measurements. The announcement says evaluation setups vary by model; its speed comparison uses a theoretical memory-bandwidth ceiling, not measured laptop throughput. We therefore do not infer a universal speed or quality lead from those charts.

The developer also identifies weaknesses: multi-turn chat, detailed constraints, factual recall and non-English output. Its card warns about repetitive reasoning, inaccurate reports of completed work and destructive file actions without confirmation. More loops do not add stored knowledge, and low reasoning effort does not fully disable reasoning.

For builders, the immediate opportunity is a local multimodal candidate that can also be evaluated through a low-priced managed route. Start with non-sensitive tasks and a copy of a repository, keep output-token and turn limits, and measure accepted results rather than tokens alone. Evaluate native and converted checkpoints separately. Retain your current model if tool behavior, memory headroom or task correctness fails your checks.

05

Pick the first evaluation path

A reader checklist, not a test AccessAllGPT performed.

Entries stay in this browser tab and are not submitted to AccessAllGPT. Blank responses are copied as [Unresolved].

Confirm thinkingmachines/inkling-2-mini, provider, price and data terms in your account; cap reasoning output and compare completed-task cost.

Pin a runtime and checkpoint, review the license and linked policy, then measure memory and latency at your actual context length.

Use disposable working copies, fixed tools, tests and turn limits. Require confirmation before deletions or irreversible actions.

Primary sources

  1. Introducing Inkling 2 MiniThinking Machines Lab · Reviewed: November 20 date, introduction, capabilities, limitations, architecture and Try Inkling 2 Mini today · Retrieved · Supports: First-party release, Tinker access, fixed-loop architecture, hardware guidance, vendor coding results and speed-chart methodology.
  2. Inkling 2 Mini model cardThinking Machines Lab on Hugging Face · Reviewed: General information, getting started, required hardware/software, properties, coding evaluations and limitations · Retrieved · Supports: Apache 2.0 declaration, linked acceptable-use policy, modalities, parameter accounting, context, checkpoint size and deployment caveats.
  3. Inkling 2 Mini repository inventoryHugging Face · Reviewed: Private/gated flags, cardData and siblings · Retrieved · Supports: Observed public ungated weight, MTP and configuration inventory, not downloaded or validated weights.
  4. Welcome Inkling 2 Mini, Cutting-edge Performance for your Local SetupHugging Face · Reviewed: November 20 date, architecture, inference support, Transformers, llama.cpp, SGLang and MLX · Retrieved · Supports: Integration-maintainer runtime instructions, per-visit state caveat, SM103 requirement and distinction between fixed- and variable-loop artifacts.
  5. Inkling 2 Mini on OpenRouterOpenRouter · Reviewed: Model discovery page and route · Retrieved · Supports: Public managed-route listing; availability is not a hands-on inference result.
  6. OpenRouter Models APIOpenRouter · Reviewed: thinkingmachines/inkling-2-mini ID, created, modalities, context and pricing · Retrieved · Supports: November 21 route creation, exact ID, context and per-token price snapshot.
  7. Exact Inkling 2 Mini provider endpointOpenRouter · Reviewed: PrimeIntellect endpoint, pricing, context, completion cap, quantization and supported parameters · Retrieved · Supports: Observed paid provider route and advertised limits and capabilities, not tested access.

Limitations

No hands-on inference, local runtime test, weight download, integrity check, memory measurement or benchmark reproduction. The OpenRouter route is paid. Tinker account access and pricing were not tested or quoted. Alternative quantized repositories and the variable-loop research artifact were not individually validated. Vendor benchmark comparisons use varying setups; theoretical speed is not observed throughput.

Disclosures

Public-source reporting, not a sponsored test or private-access claim. No vendor reviewed this briefing. AccessAllGPT Research is operated by NeuralArc; publication-wide relationships appear on the disclosures page.

Further AccessAllGPT guidance

  1. Managed API versus self-hosting
  2. Inkling 2 Small’s earlier API arrival
  3. What AI models cost
  4. How this publication is made