Agentic coding research
Evaluate an open-weight reasoning model in coding harnesses and tool-using agents with full control over prompts, runtime, and checkpoints.
Independent tool overview
Inkling-Small at a glance
Inkling-Small is Thinking Machines Lab's open-weight multimodal reasoning model for coding, agents, chat, retrieval, and text analysis. It accepts text, images, and audio and generates text, supports controllable reasoning effort, and exposes 276B total parameters while activating roughly 12B per token. The Apache 2.0 weights can be self-hosted or fine-tuned, while Tinker offers managed training and beta serverless inference. Despite the name, the full model is a very large deployment: 'Small' means smaller than the 975B-parameter Inkling, not laptop-sized.
Visit the official Inkling-Small site ↗
- Provider
- Thinking Machines Lab
- Release date
- July 30, 2026
- Architecture
- 42-layer sparse Mixture-of-Experts transformer
- Parameters
- 276B total; approximately 12B active per token
- Modalities
- Text, image, and audio input; text output
- Context
- Up to 1M tokens in the released model; 64K or 256K on current Tinker variants
- Weights and license
- BF16 and NVFP4 weights under Apache 2.0, subject to the provider's Model Acceptable Use Policy
- Availability
- Hugging Face, self-hosting, Tinker fine-tuning and Playground, plus beta Tinker inference
- Reviewed
- August 31, 2026 from official release, model card, Hugging Face, pricing, and acceptable-use sources
Overview
What Inkling-Small is
Thinking Machines released the final Inkling-Small weights on July 30, 2026 after previewing the model alongside Inkling. It is a 42-layer decoder-only Mixture-of-Experts transformer with 276B total parameters, 12B active parameters per token, hybrid local and global attention, and controllable reasoning effort.
The model natively processes text, images, and audio in one decoder and returns text. Thinking Machines lists a context window of up to one million tokens for the weights, while Tinker's managed training variants currently expose 64K or 256K contexts and its beta inference endpoint uses 256K. Product limits therefore depend on the serving path, not only the theoretical model maximum.
The full weights are available in BF16 and NVFP4 under Apache 2.0, with documented deployment paths through Transformers, vLLM, SGLang, TokenSpeed, Unsloth, and community quantizations. Open weights allow inspection, fine-tuning, and self-hosting, but do not make the model cheap to run: all 276B parameters must still be stored even though only a fraction are active for each token.
Thinking Machines reports strong reasoning, coding, tool-use, instruction-following, vision, and audio results, including 80.2% on SWE-bench Verified and 31.6% on text-only Humanity's Last Exam at its stated evaluation settings. Those numbers are useful screening evidence, not a substitute for workload-specific testing; several comparisons combine provider-reported results and internal harnesses.
The model card explicitly warns about hallucinations, imperfect instruction following, long-conversation degradation, uneven language and demographic performance, and weak factual recall relative to some alternatives. Production systems should add retrieval, validation, moderation, least-privilege tools, monitoring, and human review for consequential outputs.
Use cases
Who Inkling-Small is best for
The strongest fit depends on the job you need the product to complete, not the size of its feature list.
Custom multimodal assistants
Build systems that combine text with image or audio understanding while keeping output in a conventional text interface.
Model fine-tuning teams
Customize an Apache 2.0 checkpoint locally or through Tinker for domain behavior, graders, synthetic data, or specialized agents.
Private or controlled deployments
Operate the weights in a chosen environment when a hosted closed model does not meet governance or customization requirements.
Reasoning-cost experiments
Sweep the model's thinking-effort setting to measure the latency, token, and quality tradeoff for a defined workload.
AI infrastructure teams
Use a large sparse model when the organization already has the GPU capacity, serving expertise, observability, and safety controls to run it well.
Capabilities
Core Inkling-Small features
Sparse 276B MoE architecture
Routes each token to six of 256 experts plus two shared experts, activating about 12B parameters per token rather than the full parameter count.
Native multimodal input
Processes text, pixel-based images, and 16kHz WAV audio jointly and returns text responses.
Controllable thinking effort
Lets applications vary inference effort to trade additional reasoning tokens and cost for potential performance gains.
Long-context support
The released model supports up to one million tokens, although managed Tinker endpoints currently publish smaller 64K and 256K configurations.
Open weights
Publishes BF16 and NVFP4 checkpoints under Apache 2.0 for research, modification, fine-tuning, and compatible commercial use.
Local-serving ecosystem
Official guidance points to Transformers, vLLM, SGLang, TokenSpeed, Unsloth, and available quantizations.
Tool-use and coding behavior
The model was trained and evaluated for agentic coding, function use, multi-step instructions, and general agent workflows.
Tinker customization
Managed Tinker workflows support sampling, supervised fine-tuning, reinforcement learning, adapter export, and variable effort.
Tinker Playground
Provides an easier hosted interface for trying text, image, and audio conversations without deploying the checkpoint.
Published model card
Documents architecture, training-data categories, evaluations, safety testing, intended uses, limitations, and deployment guidance.
Process
How the Inkling-Small workflow works
Step 1
Define the workload
Specify modalities, context size, latency, throughput, tool access, accuracy target, data sensitivity, and the cost of a wrong answer.
Step 2
Choose a serving path
Compare self-hosting, a third-party provider, Tinker Playground, Tinker training, and beta Tinker inference against the required reliability and control.
Step 3
Review licenses and policy
Confirm Apache 2.0 obligations and the current Thinking Machines Model Acceptable Use Policy for the intended product, users, and jurisdiction.
Step 4
Budget the real footprint
Estimate weight storage, GPU memory, quantization, routing overhead, context cache, concurrency, network transfer, and operations; do not size infrastructure from the 12B active count alone.
Step 5
Build a representative evaluation
Test real prompts, languages, images, audio, long context, factuality, refusal behavior, tool calls, structured outputs, latency, and total cost.
Step 6
Establish baselines
Compare the same test set with smaller open models and hosted alternatives before accepting provider benchmark results as relevant to the use case.
Step 7
Tune reasoning effort
Sweep effort levels and output limits to find the smallest reasoning budget that meets the quality target without uncontrolled token growth.
Step 8
Add grounding and validation
Use current retrieval sources, schema validation, deterministic business rules, citations, and independent checks for facts or tool arguments.
Step 9
Constrain agent tools
Give tools least privilege, validate every argument outside the model, isolate untrusted content, and require human approval for consequential actions.
Step 10
Harden multimodal input
Treat uploaded files as untrusted, enforce type and size limits, scan content, strip unnecessary metadata, and guard against instructions embedded in images, audio, or documents.
Step 11
Pilot with observability
Log privacy-safe traces, model and infrastructure cost, response quality, refusal and escalation rates, tool failures, and human corrections.
Step 12
Re-test every change
Run the evaluation suite when changing checkpoints, quantization, serving software, prompts, tools, context length, fine-tunes, or safety layers.
Cost
Inkling-Small pricing and free plan
The model weights are downloadable without a model license fee, but self-hosting still carries substantial compute, storage, and operations costs. Tinker publishes per-million-token training rates and a separate beta serverless inference price. Its Inkling-Small training rates were under a limited-time 50% discount when checked, so the crossed-out standard rates are also included below and buyers should verify current billing before a run.
Open weights / self-hosted
$0 model license fee
Download Apache 2.0 BF16 or NVFP4 weights and operate them on your own or rented infrastructure.
- 276B total parameters still require substantial storage and memory
- GPU, hosting, bandwidth, engineering, monitoring, and support are separate
- Third-party quantizations and inference providers have their own terms and prices
- Use remains subject to the current model acceptable-use policy
Tinker training — 64K
Promo: $0.58 prefill / $1.44 sample / $1.73 train per 1M tokens
Managed 64K-context training access during the published limited-time 50% discount.
- Standard rates shown as $1.16 prefill, $2.88 sample, and $3.46 train per 1M tokens
- Cached prefill receives an 80% discount; promotional cached rate shown as $0.116 per 1M tokens
- Checkpoint storage is $0.10 per GB per month
- Rates and promotion can change
Tinker training — 256K
Promo: $1.16 prefill / $2.89 sample / $3.47 train per 1M tokens
Managed longer-context training during the published limited-time 50% discount.
- Standard rates shown as $2.32 prefill, $5.78 sample, and $6.94 train per 1M tokens
- Cached prefill promotional rate shown as $0.232 per 1M tokens
- Checkpoint storage is billed separately
- The managed context is 256K, not the weights' full one-million-token maximum
Tinker serverless inference beta
$0.30 input / $1.20 output per 1M tokens
A 256K-context NVFP4 hosted endpoint for Inkling-Small.
- Cached input is listed at $0.06 per 1M tokens
- The service is explicitly labeled beta
- Thinking Machines does not recommend it for intensive production use yet
- Production-volume users are directed to join a waitlist
Pricing checked . Check current pricing at the source ↗
Assessment
Inkling-Small strengths and limitations
Where it stands out
- Apache 2.0 weights provide unusually broad freedom to inspect, customize, and deploy the model
- Combines text, image, and audio understanding in a single open-weight reasoning model
- Activating about 12B parameters per token can improve inference efficiency relative to the 276B total size
- Controllable effort exposes an explicit quality, latency, and token-cost dial
- Official deployment support spans several major open-model serving stacks
- Tinker provides a managed path for fine-tuning and reinforcement learning without building the entire training stack
- The provider publishes detailed architecture, evaluation, safety, training-data category, and limitation documentation
- Provider-reported coding, reasoning, instruction-following, and multimodal results make it a credible candidate for evaluation
What to consider
- The word 'Small' is relative to Inkling: the full checkpoint has 276B parameters and remains expensive to store and serve
- Twelve billion active parameters per token does not mean the model fits like a dense 12B model; the whole expert set must remain accessible
- The model returns text only even when its inputs include images or audio
- The one-million-token model maximum is not available on every managed path; current Tinker configurations publish 64K or 256K
- Tinker's serverless inference is beta and is not recommended by the provider for intensive production use
- Promotional Tinker training prices are temporary and can materially change project economics
- Many benchmark numbers are provider-reported or use internal harnesses and may not transfer to a specific application
- The provider's own results show weaker factuality than larger Inkling and several comparison models on SimpleQA Verified
- The model can hallucinate, miss instructions, degrade across long conversations, and perform unevenly across languages or demographic groups
- Open weights shift serving security, patches, monitoring, abuse prevention, capacity planning, and incident response to the deployer
- Agent and tool use creates prompt-injection, excessive-permission, unsafe-argument, data-exfiltration, and irreversible-action risks
- Training sources are described by category rather than enumerated, so teams cannot fully trace every data origin from public documentation
- Apache 2.0 licensing coexists with a separate Model Acceptable Use Policy that deployers must review and enforce
- Medical, legal, financial, biometric, and other consequential uses require domain validation and meaningful human oversight; the model is not an autonomous decision authority
Compare
Inkling-Small alternatives
The right alternative depends on the specific output, workflow, controls and budget your project requires.
Consumer
Kimi K2.6
A larger open-weight MoE alternative with strong agentic and vision positioning for teams comparing frontier-scale customizable models.
Explore Kimi K2.6 →Coding
Qwen3-Coder-Next
A more coding-focused open-weight model for teams that prioritize agentic software tasks over audio capability.
Explore Qwen3-Coder-Next →Consumer
Gemma 4
Google's open-weight family offers multiple sizes for teams seeking a broader deployment-footprint range.
Explore Gemma 4 →Business Operations
DeepSeek
A widely deployed model and hosted assistant ecosystem for teams comparing cost, reasoning, and self-hosted options.
Explore DeepSeek →Coding
Ollama
A local-model runtime rather than a model, useful when the main priority is easy operation of smaller compatible checkpoints.
Explore Ollama →Questions
Inkling-Small FAQs
What is Inkling-Small?
Inkling-Small is Thinking Machines Lab's open-weight multimodal reasoning model. It accepts text, images, and audio, produces text, supports tool use and variable reasoning effort, and can be self-hosted or customized through Tinker.
Is Inkling-Small open source?
Its model weights are released under the permissive Apache 2.0 license, so 'open-weight' is the most precise description. The provider also publishes a Model Acceptable Use Policy that users and deployers must follow.
How large is Inkling-Small?
It has 276B total parameters and activates about 12B per token through sparse Mixture-of-Experts routing. It is one quarter of the total parameter count of the larger Inkling model, but it is still a very large checkpoint.
Can Inkling-Small run on a laptop?
The official full-precision and NVFP4 checkpoints are designed for substantial accelerator infrastructure. Community quantizations may broaden hardware options, but speed, memory use, context length, multimodal support, and quality should be tested rather than inferred from the 12B active count.
What inputs and outputs does Inkling-Small support?
It accepts UTF-8 text, pixel-based images, and 16kHz WAV audio and generates UTF-8 text. The model card recommends image dimensions from 40 to 4096 pixels and audio clips ideally under two minutes.
Does Inkling-Small support one million tokens?
The released model is described as supporting up to a one-million-token context. Current Tinker training variants expose 64K or 256K, and beta Tinker inference publishes 256K, so verify the limit for the exact runtime.
How much does Inkling-Small cost?
The weights have no model license fee, but self-hosting compute and operations can be substantial. On August 31, 2026, Tinker listed discounted training rates and beta hosted inference at $0.30 per million input tokens, $0.06 cached input, and $1.20 per million output tokens.
Is Tinker inference ready for production?
Thinking Machines labels serverless inference beta and says it does not recommend the service for intensive production use yet. Teams with production volume or latency needs are directed to a waitlist.
How good is Inkling-Small for coding?
Thinking Machines reports 80.2% on SWE-bench Verified, 55.9% on SWE-bench Pro public, and 64.7% on Terminal-Bench 2.1 using the stated harnesses and settings. Run a private repository evaluation because benchmark conditions, tool access, contamination controls, latency, and task mix can differ from production.
Is Inkling-Small more capable than Inkling?
It exceeds the larger model on some published reasoning, coding, and instruction-following evaluations while using less active compute, but Inkling retains stronger knowledge coverage and factuality on the provider's published results. Neither is uniformly better.
Can Inkling-Small be fine-tuned?
Yes. The Apache 2.0 weights can be customized with compatible infrastructure, and Thinking Machines offers managed supervised and reinforcement-learning workflows through Tinker.
What safety controls are needed?
Use application-layer moderation, retrieval and fact checks, least-privilege tools, deterministic argument validation, rate limits, privacy-aware logging, adversarial tests, human approval for impactful actions, and a kill switch. Do not rely on model refusals alone.
Bottom line
Our Inkling-Small verdict
Inkling-Small is an attractive open-weight candidate for teams that need strong reasoning, agentic coding, and native image or audio understanding with the freedom to customize the model. Its main trap is the name: 12B active parameters improve sparse inference economics, but a 276B checkpoint is still infrastructure-heavy. Shortlist it only after measuring the exact serving footprint, factuality, safety, latency, and cost against smaller open models and stable hosted APIs on your own workload.
Visit Inkling-Small website ↗Similar Miscellaneous Tools
View all Miscellaneous AI tools →.jpeg)
AI training for the future of work.
Get access to all our AI courses, hundreds of real-world AI use cases, live expert-led workshops, an exclusive network of AI early adopters, and more.
AI Courses
Get unlimited access to all of our current & upcoming industry-specific AI courses for the duration of your subscription.
Daily Guides
To keep up with the rapid pace of AI, our team publishes AI implementation guides daily. Our library contains 300+ practical use cases to automate real-world work.
Workshops
Join weekly, live, interactive sessions with industry leaders who are at the forefront of AI for hands-on implementation guidance and exclusive insights.
Community
Network with an exclusive community of AI-first professionals who are working smarter with AI. Learn how early adopters are using AI in their work and businesses.