Inkling 2 Small: Thinking Machines ships a 594B open-weights MoE that runs on 17B active parameters
Thinking Machines Lab released Inkling 2 Small with open weights under Apache 2.0: a natively multimodal 594B mixture-of-experts model with 17B active parameters, 1M tokens of context, and agentic-coding benchmarks that beat most open flash-class rivals. Self-hosting needs a GPU cluster; Ollama Cloud offers it at 0.30 dollars per million input tokens.
Thinking Machines Lab released Inkling 2 Small this week with open weights under Apache 2.0, and the name is doing some work: the model is a 594B-parameter mixture-of-experts where only 17B parameters are active per token. It is natively multimodal, reads text and images, supports up to 1M tokens of context, and is built for agentic coding, research, and knowledge work. Within two days of release it was on Hugging Face, on Thinking Machines' own Tinker platform, and on Ollama Cloud at 0.30 dollars per million input tokens and 1.00 dollar per million output tokens, with cached input at 0.03.
The short version for builders: this is a serious open-weights option in the flash class, trained natively in NVFP4, with agentic-coding numbers that beat most open rivals on the vendor's own table. But "Small" does not mean laptop-friendly: self-hosting the weights needs a GPU cluster, so for most people the practical entry point is a hosted API.
What the model actually is
Per the model card, Inkling 2 Small is a 48-layer decoder-only transformer with a sparse MoE feed-forward backbone: each token routes to 8 of 384 experts. Attention is a hybrid of local and global layers. Images are encoded by a vision transformer and processed jointly with text in a shared hidden space, so vision is native rather than bolted on. It takes text in UTF-8 and images in any pixel-based format, and generates text.
Two numerics details matter for deployment. The model was pretrained, midtrained, and post-trained in NVFP4, which is its native precision, and the vendor recommends that checkpoint for all use cases. An alternative MXFP8 checkpoint gains a marginal 1 to 2 percentage points on evals, which Thinking Machines itself says is not meaningful for practical use.
The benchmarks, all vendor-reported
Thinking Machines reports 86.3% on SWE-bench Pro v2 (Hard), 72.1% on DeepSWE v1.1, 85.4 on MCP Atlas (a tool-use eval), and 81.6% on MMMU Pro for multimodal understanding. On its own comparison table against open and closed flash-class models, Inkling 2 Small leads the open-weights flash group on most rows: DeepSWE 72.1% versus 63.4% for GLM-5.3-Flash, though DeepSeek-V4.1-Flash-1017 edges it at 77.0%. On the third-party Artificial Analysis index (v5.0) it scores 30.6, behind DeepSeek-V4.1-Flash at 32.6 and Claude Haiku 5.5 at 36.5, ahead of GLM-5.3-Flash at 27.8.
Read these as positioning, not verdicts. The table is the vendor's, the eval versions are specific (SWE-bench Pro v2, DeepSWE v1.1 with mini-swe-agent v2.4.6), and I did not run any of them.
How to actually run it
Four paths, per the model card and the Ollama listing:
- Hugging Face: open weights in two checkpoints (NVFP4 main, MXFP8 alternative), Apache 2.0.
- Tinker: Thinking Machines' fine-tuning platform, API access plus an inference-only playground.
- Third-party inference providers, per the vendor's partner docs.
- Ollama Cloud: ollama run inkling-2-small:cloud, with ready-made launch commands for Claude Code, OpenCode, Hermes Agent, and OpenClaw, at 0.30 / 0.03 cached / 1.00 dollars per million tokens.
Self-hosting is the catch. The NVFP4 checkpoint requires at least 500GB of aggregated VRAM, for example 2x NVIDIA B300 or 4x H200 GPUs. The MXFP8 checkpoint needs 800GB, so 4x B300 or 8x H200. Running it also requires a serving framework: SGLang, vLLM, TokenSpeed, Unsloth, or Hugging Face Transformers. In other words, the weights are open in the license sense, not in the runs-on-your-machine sense.
The safety section is worth reading
The model card's safety work is unusually specific for an open release: multi-turn external red-teaming on sycophancy, harmful manipulation, and parasocial-dependency patterns; CBRN and cyber uplift assessed with refusal-suppressed variants to estimate latent capability; and loss-of-control evals covering agentic capability, strategic deception, and sabotage potential, where the vendor reports the model materially below frontier capabilities. Their conclusion: no material uplift beyond what already exists in the open-weight ecosystem.
They still recommend human oversight for high-stakes outputs, application-layer safeguards like content filtering and rate limiting, and avoiding medical, legal, or safety-critical deployment without domain-specific fine-tuning and validation.
What I would check
My untested suggestions before adopting it:
- Try the Ollama Cloud route first for a day of real work; at these prices the experiment is nearly free, and the Claude Code launch command makes it a one-line swap.
- If you care about tool use, weight MCP Atlas (85.4) more than generic chat benchmarks; that is the number closest to agent workloads in the vendor's table.
- Verify the NVFP4 quality story on your own tasks before assuming the cheaper checkpoint is equivalent; the vendor says the MXFP8 gain is meaningless, but your eval is the one that counts.
- If compliance matters in your organization, read the training-data section: sources include public internet, third-party acquisitions, internal, and synthetic data, with dedup and filtering described but no dataset manifest.
- Budget for the cluster reality before promising anyone a self-hosted deployment; 500GB VRAM is a procurement conversation, not a weekend project.
What I would not conclude
Do not conclude "Small" means small; 594B total parameters is a large model with a sparse activation pattern, and self-hosting requirements match that. Do not treat the vendor benchmark table as neutral; it is well-organized but still the vendor's selection of evals and rivals. Do not assume open weights means unrestricted use in every context; the license is Apache 2.0, but the vendor's own deployment guidance adds oversight expectations for high-stakes domains. And do not read the safety conclusion as an external audit; the red-teaming was commissioned by the vendor.
Frequently asked questions
What is Inkling 2 Small?
A general-purpose multimodal model from Thinking Machines Lab: 594B total parameters in a mixture-of-experts design with 17B active per token, 1M tokens of context, text and image input, released with open weights under Apache 2.0.
Is Inkling 2 Small really open source?
The weights are downloadable from Hugging Face under the Apache 2.0 license, which permits commercial use, fine-tuning, and redistribution. Training data and code are not released, so it is open weights rather than fully open source.
Can I run Inkling 2 Small locally?
Only with serious hardware: the native NVFP4 checkpoint needs at least 500GB of aggregated VRAM (for example 2x B300 or 4x H200 GPUs), and the MXFP8 variant needs 800GB. Most developers will use a hosted API instead.
How much does the hosted version cost?
On Ollama Cloud it is listed at 0.30 dollars per million input tokens, 0.03 for cached input, and 1.00 dollar per million output tokens. It is also available through Thinking Machines' Tinker platform and third-party inference providers.
How good is it at coding?
Per the vendor: 86.3% on SWE-bench Pro v2 (Hard) and 72.1% on DeepSWE v1.1, competitive within the flash class but behind DeepSeek-V4.1-Flash on some rows. These are vendor-reported numbers on specific eval versions; independent verification was not available at release time.
Does it support vision?
Yes, natively. Images are encoded with a vision transformer and processed jointly with text. The vendor lists document reading, charts, screenshots, OCR, and visual reasoning, with 81.6% on MMMU Pro.
What did the safety evaluation find?
The vendor reports evaluations for sycophancy and manipulation patterns, CBRN and cyber uplift, and loss-of-control risks, concluding the model is materially below frontier capabilities and adds no material uplift beyond the existing open-weight ecosystem. They still recommend human oversight and application-layer safeguards for deployment.
Sources
- Inkling 2 Small Model Card | Thinking Machines Lab - read November 15, 2026. Architecture, parameters, context, training numerics, hardware requirements, distribution, the full eval table, and the safety evaluation.
- inkling-2-small | Ollama - read November 15, 2026. Cloud pricing, context, the launch commands for coding agents, and the vendor benchmark summary.
- thinkingmachines/Inkling-2-Small | Hugging Face - read November 15, 2026. License confirmation (Apache 2.0), safetensors distribution, and serving-framework snippets.
One secondary report claimed the weights were closed; the vendor model card and the Hugging Face listing both confirm open weights under Apache 2.0, so that claim is wrong. All benchmark figures are vendor-reported; I did not run the model. A zero-data-retention claim circulating in newsletter coverage could not be confirmed from the primary sources and is omitted.
Related reading
- LongCat-2.5: Meituan open-sources a 1.6T agent model under MIT, trained and served without Nvidia
Meituan's LongCat team released the weights for LongCat-2.5, a 1.6-trillion-parameter mixture-of-experts model with about 48B active parameters per token, native image understanding, a 512K-token context, and an MIT license. Training and the reference deployment run entirely on domestic Chinese accelerators, with W4A4 INT4 quantization that the team says roughly matches full-precision accuracy. Vendor benchmarks put it near Claude Opus 5 on agentic coding, below it on terminal tasks. Self-hosting needs a serious cluster.
- Gemini 4 Hydrogen: Google's fast execution model undercuts Argon on price and beats it on intelligence
Google launched Gemini 4 Hydrogen on November 12: a fast everyday coding and agentic model that scores 51 on the Artificial Analysis Intelligence Index at its default thinking level, above Gemini 4 Argon's 50, at 1.50 dollars per million input tokens and 7.50 per million output tokens. What the three-model Gemini 4 ladder means, how Orchestrator splits work in Antigravity, and what to verify.
- Gemini 4 Carbon GA: what paying developers get today, and the November 30 price jump
Google made Gemini 4 Carbon generally available on November 10: paid API access in AI Studio and Gemini for business, a /swarm multi-agent preview in Antigravity, Fairwind cybersecurity access, and introductory pricing that doubles after November 30, 2026. What changed since the October announcement, what Google claims, and what to verify yourself.
The build log
Follow the work.
New essays and prototypes, in your feed reader. Email subscriptions are paused while I connect a working delivery service. No email is collected here.
Follow via RSS