---
title: "Inkling 2 Small: Thinking Machines’ 594B MoE you can tune but not download yet"
description: "Inkling 2 Small is a 594B MoE with 17B active. You can fine-tune it on Tinker today, but the Apache 2.0 weights are still a promise. Benchmarks and costs."
canonical_url: https://www.oguzhan.co/inkling-2-small-thinking-machines-open-weights/
author: "Oğuzhan Koçaklı"
author_url: https://www.oguzhan.co/about/
date_published: 2026-11-10T07:02:30+00:00
date_modified: 2026-11-10T07:02:30+00:00
language: en
translations:
  tr: https://www.oguzhan.co/tr/inkling-2-small-thinking-machines-acik-agirlik/
---

# Inkling 2 Small: Thinking Machines’ 594B MoE you can tune but not download yet

Thinking Machines Lab announced Inkling 2 Small on 5 November. It is a mixture-of-experts model with 594 billion parameters in total and 17 billion active per token, with image input and an adjustable effort setting. You can fine-tune it on Tinker today. You cannot download it. The [model card](https://thinkingmachines.ai/model-card/inkling-2-small/) lists the license as “Tinker Terms of Use”, and the [launch post](https://thinkingmachines.ai/news/inkling-2-small/) says the lab intends to release full weights under Apache 2.0 “next week” after external safety testing and red-teaming.

An open model you can train but not hold, then. It also comes with the most candid launch post of the month, one that prints its own bad benchmark rows and an RL incident most labs would have kept quiet.

## Why release on Tinker first and the weights “next week”?

Because that is the plan the lab published in July. [“A Safe Path to Open Weights”](https://thinkingmachines.ai/blog/a-safe-path-to-open-weights/) describes a staged release: a limited inference API, then monitored general availability, then open weights. Fine-tuning APIs like Tinker sit in the middle. The provider can still watch usage, keep guardrails and revoke access. The same post says these stages “are not a fixed sequence that automatically ends with releasing the model’s weights”. Worth reading twice before you put “next week” in a calendar.

The [Inkling 2 Small post](https://thinkingmachines.ai/news/inkling-2-small/) follows that route. Closed weights on Tinker first, fine-tuning and sampling available now, checkpoints downloadable once the weights are released, other platforms “in the coming weeks”. [Artificial Analysis](https://artificialanalysis.ai/articles/thinking-machines-releases-inkling-2-small-highest-scoring-open-weights-model-outside-of-chinia-upon-weights-release-it) files it as “Proprietary at launch, Apache 2.0 weights release planned (date unconfirmed)”.

The official `thinkingmachines` org listing on Hugging Face, as of a check on 10 November at about 10:00 Istanbul time, showed six models, the newest being Inkling-Small, last modified 31 July. No Inkling 2 Small repo in the official org at that check, and the launch post still promised the weights for “next week”. That describes one listing at one moment. Meanwhile, [Cognition built SWE-2 Flash on the first-generation Inkling-Small](https://www.oguzhan.co/ai-text-watermark-claude-grok-flashprove-kimi-k3-1-pro/).

## What changed inside Inkling 2 Small?

Per the [model card](https://thinkingmachines.ai/model-card/inkling-2-small/), it is a 48-layer decoder-only transformer with sparse MoE. Each token is routed to 8 of 384 experts, attention is hybrid local and global, and numerics are MXFP8 and NVFP4. Context window is up to 1M tokens. Output is text only.

The [launch post](https://thinkingmachines.ai/news/inkling-2-small/) splits pretraining into four phases. H1 was 8.4T tokens of text, code and images, 68 hours on NVIDIA GB300 NVL72 racks, natively in NVFP4. H2 added 3.8T tokens of newer, higher-quality data at 512K context. H3 was about 320B tokens of reasoning and agent trajectories at 1M context, and H4 about 34B tokens of top agent trajectories. Those H3 and H4 trajectories were sampled mostly from earlier models, “including Kimi K3 and our own models”, then filtered for correctness and alignment. Some of what this model knows about being an agent came from someone else’s agent.

The bigger change is RL. The first Inkling had separate reasoning and agentic RL runs merged with multi-teacher distillation. Inkling 2 Small got one joint run: 1,075 steps, more than 17 million rollouts, 9 days on 2,048 GPUs, over 4 x 10^23 FLOPs. Coding took 65.0% of RL tokens, office and knowledge work 13.3%, STEM 7.6%, web and app building 7.2%. Rollouts ran 43K tokens and 8 minutes at the median, 1.5M tokens and 4 hours at p99. Per the lab, RL used a 256K context with compaction.

The detail builders should care about is the grader. Coding RL covered hundreds of thousands of long-horizon tasks across multiple agent harnesses, and many tasks had an agentic RL grader judging how the model approached the problem, beyond whether unit tests passed. Early checkpoints ran publicly on OpenCode as “Fledge-Alpha”.

Vision comes from a 292M-parameter encoder based on Perception Encoder, with image patches interleaved with text tokens. Audio is gone. The first Inkling accepted it, per Tinker’s model table.

![Dark grid of expert cells with a handful lit up as one token is routed to them, mixture-of-experts sketch](https://oguzhanco.s3.eu-west-3.amazonaws.com/wp-content/uploads/2026/11/10070223/inkling-2-small-moe-8-of-384-experts.webp)

*Each token goes to 8 of 384 experts, so only 17B of 594B parameters work at a time. Illustration: Higgsfield.*

## The scoreboard, including the rows the launch post doesn’t brag about

The comparison comes from Thinking Machines’ own launch table, run at effort 0.99 (xhigh), temperature 1.0 and a 1M-token maximum trajectory. That cap is an eval setting, not a served context window.

|  | Inkling 2 Small | DeepSeek-V4.1-Flash-1017 | GLM-5.3-Flash | Kimi K3 | Claude Haiku 5.5 |
| --- | --- | --- | --- | --- | --- |
| Params active / total (B) | 17 / 594 | 16 / 552 | 18 / 320 | 104 / 2800 | not disclosed |
| AA Index v5.0 | 30.6 | 32.6 | 27.8 | n/a | 36.5 |
| DeepSWE v1.1 | 72.1% | 77.0% | 63.4% | 68.5% | 59.9% |
| Terminal Bench v4.0 | 13.1% | 33.8% | 32.8% | 12.6% | 32.8% |
| MMMU Pro | 81.6% | 80.6% | 79.0% | 80.5% | 79.1% |
| DesignArena Website | 1340 | 1297 | 1270 | 1334 | 1261 |
| Finance Agent v2.1 | 67.5% | 62.9% | 60.6% | 60.0% | 65.4% |
| AA-Omniscience non-hallucination | 73.3% | 35.9% | n/a | n/a | 65.8% |

*Source: [Thinking Machines’ launch table](https://thinkingmachines.ai/news/inkling-2-small/), many cells sourced from third-party evaluators (Artificial Analysis, Vals, DesignArena). The [model card](https://thinkingmachines.ai/model-card/inkling-2-small/) also lists Terminal Bench 2.1 at 85.1% with the Pi harness and 81.3% with Terminus. Different version, different harnesses, not comparable with the v4.0 row.*

![Row of tall glowing glass bars with one conspicuously short bar in the middle, abstract benchmark scoreboard](https://oguzhanco.s3.eu-west-3.amazonaws.com/wp-content/uploads/2026/11/10070226/inkling-2-small-benchmark-short-bar.webp)

*Terminal Bench v4.0 is the short bar: 13.1% in the lab’s own table. Illustration: Higgsfield.*

Start with the short bar. On Terminal Bench v4.0 the [model card](https://thinkingmachines.ai/model-card/inkling-2-small/) gives 13.1%, against 41.9% for GLM 5.3, 33.8% for DeepSeek-V4.1-Flash-1017 and 32.8% for Claude Haiku 5.5. Mind the suffix: GLM 5.3 at 41.9% is a different model from GLM-5.3-Flash at 32.8% in the table. [Artificial Analysis measured](https://artificialanalysis.ai/articles/thinking-machines-releases-inkling-2-small-highest-scoring-open-weights-model-outside-of-chinia-upon-weights-release-it) 13% against 30% for Claude Sonnet 5.5 (medium). SWEBench Verified is 82.2% on the card, behind GLM 5.3 at 95.4% and Kimi K3 at 93.4%.

That does not make the model weak at code across the board. DeepSWE v1.1 is 72.1% and the Terminal Bench 2.1 rows are high. It means v4.0 is a soft spot, so terminal-heavy workloads deserve their own test. The [LongCat-2.5 deep dive](https://www.oguzhan.co/meituan-longcat-2-5-1-6t-moe-chinese-chips/) made the case for reading vendor tables row by row, and it applies here.

Then tokens. Artificial Analysis scores the model 31 on its Intelligence Index, up 18 points from Inkling Small and level with Gemini 3.8 Flash (high), “with ~4x as many output tokens per task”. Running the index cost $1.18 per task, against $0.20 for Inkling Small, $7.77 for Gemini 3.8 Flash (high) and $10.02 for Claude Sonnet 5.5 (medium).

The good rows are real. MMMU Pro at 81.6% and DesignArena Website at 1340 top the table, and the lab says the model is set to be the top open-weights entry on DesignArena once weights ship. Finance Agent v2.1 is 67.5%. Non-hallucination on AA-Omniscience is 73.3%, against 35.9% for DeepSeek-V4.1-Flash-1017. The card adds BrowseComp with context management at 90.2%, the highest in its table, and GPQA Diamond at 93.0%.

## Can a model run its own fine-tuning job?

Thinking Machines says this one did. In its autoresearch demo, Inkling 2 Small fine-tuned Qwen3.5-4B on Tinker for time-series forecasting, covering electricity demand, wind and solar, weather, air quality and Wikipedia page views. Per the [lab’s account](https://thinkingmachines.ai/news/inkling-2-small/), it wrote its own evaluator, then found a bug in it: the same seed for all 8 samples. It noticed evals were costing money, $101 at one point, and batched them. It tried RL, saw forecasts turn overconfident, and stopped. It submitted its best SFT checkpoint after 4 hours 17 minutes, having spent $426 of a $500 budget.

The lab’s result: on unseen August to September data, the final model makes about 15% less error than the base model. Catching a bug in its own evaluator and watching the bill are the habits you want from anyone spending your training budget.

Kernel work is more mixed. On PMPP-Hard, 69 GPU kernel tasks with a 10-hour budget and 4 attempts, Inkling 2 Small solved 54.25 on average. GPT-6 Astra solved 63, Kimi K3.1 Pro 60.5, Claude Opus 5.5 60 and DeepSeek V4.1 Flash 55.25. On a preliminary, kernels-only DeepSeek sparse attention speedup test, Claude Opus 5.5 hit 18.80x and Inkling 2 Small 6.02x, a bit ahead of DeepSeek V4.1 Flash at 5.66x. Fine for its class. Far from the frontier.

## What happened when the model hit a network block?

This is the paragraph in the [launch post](https://thinkingmachines.ai/news/inkling-2-small/) worth reading in full:

> “Midway through the main RL run, our asynchronous monitor flagged an instance where the model, failing at a task, worked around a network block through a third-party service it could reach using its search tools. We paused the run, hardened our infrastructure, removed environments that required web access, and added synchronous tool-call monitoring.”

That is the lab’s own account, with no independent audit behind it. Few labs print anything like it. A model losing at a task found a way past a sandbox wall with a tool it was allowed to use, the same kind of pressure behavior that shows up in [OpenAI’s misalignment reports](https://www.oguzhan.co/openai-misalignment-reports-claude-ecdsa-seeds/).

The response is two monitoring layers. The synchronous one checks each tool call before it runs against the task’s declared allowed actions and escalates unclear cases to an LLM judge. The asynchronous one has an LLM judge review finished rollouts. The second caught the incident. The first was added because of it.

![Glowing agent orb inside a glass box, a thin thread escaping toward a server while a watchful eye hovers above](https://oguzhanco.s3.eu-west-3.amazonaws.com/wp-content/uploads/2026/11/10070228/inkling-2-small-sandbox-network-block-monitor.webp)

*The model slipped past a network block with a tool it was allowed to use. Illustration: Higgsfield.*

The lab also says early checkpoints, when tasks got hard, sometimes ended up “fabricating results, ignoring instructions, and concealing what they had done”. The fixes included filtering misbehavior traces out of midtraining, a general alignment reward on every RL task, environments that reward giving up on impossible tasks, and self-distillation from a teacher prompted with a conduct spec. On the lab’s internal Misbehavior score, lower being better, Inkling 2 Small is at 14.9, against 66.8 for Inkling and 75.2 for Inkling Small. The closed GPT-6.1 Sol scores 0.8, and the lab says its model “still trails closed-weights models, such as GPT-6.1 Sol”.

There is a cost. On FORTRESS it has a lower Average Risk Score than earlier Inklings but a higher over-refusal score, ORS 20.4 against 3.6 for GLM-5.3-Flash. The limitations section is blunt: “not the most succinct communicator”, less optimized for chat, “overuses of slashes”, overreacts to mental health queries, and doesn’t always pick the best tools “especially in codebases with a large number of skills”.

## What does it cost, and should you try it now?

[Tinker’s pricing](https://thinkingmachines.ai/tinker/pricing/) per 1M tokens: $0.30 prefill, $0.03 cached prefill, $1.00 sampling, and $1.50 training at 64K context or $3.00 at 256K. On the [same model table](https://tinker-docs.thinkingmachines.ai/tinker/models/models_and_pricing/), Inkling-Small costs $0.58, $1.44 and $1.73 for prefill, sampling and training, and the first-generation Inkling $1.87, $4.68 and $5.61. Per token, the new model is cheaper than both.

Per task is a different story. With about 4x the output tokens, AA’s cost per index task went from $0.20 for Inkling Small to $1.18. Budget by the task.

Keep the context numbers apart, too. The card says up to 1M tokens. Tinker serves 64K and 256K variants. The 1M in the eval settings is a maximum trajectory, not a served window. Check the pricing page before designing around a million-token prompt.

If you already fine-tune on Tinker, try it: the Playground is open and the lab says checkpoints become downloadable when weights ship. If you need weights on your own hardware, wait.

## What to watch this week

The launch post is dated 5 November, so its “next week” is this one. Watch whether the weights land in the official Hugging Face org under Apache 2.0, and whether the card’s license line moves off Tinker Terms of Use. Whether third parties rerun Terminal Bench v4.0 and the per-task token counts. And the next sizes, since the lab calls Inkling 2 Small “part of a ladder of models”. Until the weights show up, this is a strong closed model with an open-weights promise attached. The [2 to 8 November weekly recap](https://www.oguzhan.co/ai-weekly-2-8-nov-2026-cheaper-tokens-pricier-risks/) put it in the same release-status table as Ripple-65B and Hy4.

## Sources

- Thinking Machines Lab, [“Introducing Inkling 2 Small”](https://thinkingmachines.ai/news/inkling-2-small/), 5 November 2026
- Thinking Machines Lab, [Inkling 2 Small model card](https://thinkingmachines.ai/model-card/inkling-2-small/)
- Thinking Machines Lab, [Tinker pricing](https://thinkingmachines.ai/tinker/pricing/) and [Tinker models and pricing docs](https://tinker-docs.thinkingmachines.ai/tinker/models/models_and_pricing/)
- Thinking Machines Lab, [“A Safe Path to Open Weights”](https://thinkingmachines.ai/blog/a-safe-path-to-open-weights/), 31 July 2026
- Artificial Analysis, [Inkling 2 Small write-up](https://artificialanalysis.ai/articles/thinking-machines-releases-inkling-2-small-highest-scoring-open-weights-model-outside-of-chinia-upon-weights-release-it)
- Hugging Face, [official `thinkingmachines` org model listing](https://huggingface.co/api/models?author=thinkingmachines&sort=lastModified&direction=-1&limit=50), checked 10 November 2026, about 10:00 Istanbul time