Tasks and inputs
Inkling 2 Small is listed for Image-text-to-text, Text generation, Coding & Code Generation. It accepts text, image and produces text. Use the publisher’s input format and preprocessing for your chosen deployment.
SyncDev model directory · Profile updated Nov 17, 2026
Inkling 2 Small is Thinking Machines' open-weight 594B-A17B MoE with image input, adjustable effort and MCP tool use, served by many hosts at low cost.
The specification score summarizes recorded features and limits. It is not a benchmark result.
Pricing conditions: No first-party per-token price. Hosts list about $0.30 in / $1.00 out per 1M tokens (Cloudflare adds $0.10 cached); Prime Intellect $0.15/$0.50, DeepInfra $0.285/$0.95. Indicative; confirm on your provider account.
Recorded modality support shows what this model can accept and produce. An undocumented modality is shown as unknown rather than assumed to be unsupported.
Accepts
Produces
Seventeen billion active parameters doing the work of a frontier-class agent: that is the pitch for Inkling 2 Small, the second model generation from Thinking Machines Lab. It is a sparse mixture-of-experts with 594B total parameters, it reads images as well as text, and since 12 November 2026 its weights are public under Apache 2.0.
The launch on 5 November was closed-weight, on the lab's Tinker platform only. A week later the checkpoints arrived, which is why some launch-week descriptions still call the weights pending. They are not.
| Field | Value |
|---|---|
| API ID (aggregators) | thinkingmachines/inkling-2-small |
| Architecture | MoE, 48 layers, 8 of 384 experts per token, mixed local/global attention |
| Parameters | 594B total / 17B active |
| Input | Text, image (about 292M-parameter vision encoder); no audio |
| Output | Text |
| Context | Up to 1,048,576 tokens (open-weights listing and hosts) |
| Precision | Trained natively in NVFP4; MXFP8 checkpoint also published |
| Licence | Apache 2.0 |
| Indicative price | $0.30 in / $1.00 out per 1M tokens on most hosts |
The open-weights listing and hosted providers give 1M tokens. The lab's technical notes never state a serving window; it mentions a 512K-context training phase and later agent trajectories at up to 1M. Some catalogues also list a 131,072-token output cap that Thinking Machines itself does not state. Set max_tokens explicitly on your provider and read the cap back from the response before you plan around it.
The NVFP4 checkpoint wants roughly 500GB of aggregate VRAM, for example two B300s or four H200s. MXFP8 needs about 800GB and, by the lab's account, buys only one or two eval points. Serving recipes exist for SGLang, vLLM, TokenSpeed and Unsloth. For most teams the hosted route is simpler: NVIDIA, Fireworks, Together, Baseten, Nebius, Modal and Cloudflare all serve it. Prices vary. Cloudflare charges $0.30/$1.00 with $0.10 cached, Prime Intellect lists $0.15/$0.50 and DeepInfra $0.285/$0.95. Treat all of these as indicative and check your own account.
Against Inkling it scores higher on the Artificial Analysis Intelligence Index (about 31) and follows instructions better, partly because reasoning and agentic training happened in one joint RL run rather than separate runs plus distillation. Against Inkling Small, the lab reports far fewer misbehaviour cases on its internal alignment eval (14.9 versus 75.2, lower is better). Publisher-reported scores include SWE-Bench Pro v2 (Hard) at 86.3%, ARC-AGI-2 at 81.7% and MMMU Pro at 81.6%. It does not lead every row.
Known weak spots, admitted by the lab: answers are less concise, chat formatting is weaker, and in codebases with many skills it sometimes picks the wrong tool.
More from the lab sits on the Thinking Machines provider page.
Compare the model on the work you need to complete.
Inkling 2 Small is listed for Image-text-to-text, Text generation, Coding & Code Generation. It accepts text, image and produces text. Use the publisher’s input format and preprocessing for your chosen deployment.
Documented limits: 1.05M context. The listed route has starting rates of $0.30 input and $1.00 output per million tokens. Measure a completed task using the prompts, documents and tools your application will use. Include reasoning and retries in the result.
The weights are downloadable under Apache-2.0. Check runtime support and size the complete model and working context for your expected concurrency.
These records share a directory category. Check the source card, output type and deployment requirements before treating a model as a replacement.
Microsoft
Aion 1.0 Instruct 2.3B is Microsoft's MIT-licensed on-device small language model and Phi Silica successor for Windows, with a 32K context.
Gemini 4 Neon is Google's mid-tier Gemini 4 model for complex coding and analysis, sitting between Hydrogen and Carbon at $3/$15 per 1M tokens.
OpenAI
GPT-6.2 Astra is OpenAI's November 2026 flagship for hard professional, scientific and agentic work, at $8/$40 per million tokens with a 1.05M context.
OpenAI
GPT-6.2 Sol is OpenAI's everyday model for coding, agents and writing: $2/$10 per million tokens, 1.05M context, released 17 November 2026.
Alibaba
Qwen4 Max (qwen4-max) is Alibaba's first Qwen4-generation flagship: 1M-token context, text, image and video input, thinking on by default, $1.50/$4.50 per 1M.
Sarvam
Sarvam-105B-1711 is a refreshed open post-train of Sarvam's 105B MoE, tuned for tool calling, instruction following, agents and Indian languages.