Thinking Machines releases Inkling 2 Small, a 594B MoE with 1M context
Thinking Machines Lab released Inkling 2 Small on November 5, 2026, a sparse Mixture-of-Experts model with 594 billion total parameters and 17 billion active per token, a context window of up to 1 million tokens, and text plus image input. The model is available today only through Tinker, the company's fine-tuning and inference platform, where serverless inference costs $0.30 per million input tokens, $0.03 per million cached input tokens and $1.00 per million output tokens.
The release is the second generation of the "Small" line that the lab opened in July with Inkling-Small, a 276B-parameter model with 12B active parameters whose weights were published at launch. This time the order is reversed. Inkling 2 Small ships as a closed-weights model, and Thinking Machines says it intends to release the full weights under an Apache 2.0 license the following week, once external safety testing and red-teaming by the company and its partners are complete. Until that happens, developers can sample from the model and fine-tune it on Tinker, but they cannot download or self-host it.
On the Artificial Analysis Intelligence Index v5.0, the company reports a score of 30.6 for Inkling 2 Small, which it rounds to 31 in the announcement. Its strongest published results are in browsing, screen understanding and web design tasks, while several coding and reasoning rows in its own comparison table are led by other open-weight models such as DeepSeek-V4.1-Flash and Kimi K3.
··········
INKLING 2 SMALL ARCHITECTURE, CONTEXT AND MODALITIES.
According to the model card, Inkling 2 Small is a 48-layer decoder-only transformer with a sparse MoE feed-forward backbone. Each token is routed to 8 of 384 experts, and the attention stack mixes local and global layers. The model accepts text and images and produces text. Vision runs through a 292M-parameter encoder based on Perception Encoder, with 20×20-pixel patches and 2×2 spatial pooling, and the card recommends images between 40 and 4,096 pixels per side for best results.
The 1M-token context is a property of the model, and the training schedule was built around it. The lab says benchmark trajectories were run with a 1M-token limit, up from 256K in the evaluations of Inkling-Small, while reinforcement learning used a 256K window with compaction for longer runs. On Tinker, the price list currently exposes the model at 64K and 256K context tiers, so the full 1M window described in the model card is not what a developer gets by default on the platform today.
Inference supports variable reasoning effort. The published evaluations were run at an effort setting of 0.99, labelled "xhigh", with temperature 1.0, which is the configuration readers should assume behind every score quoted below.
........
Specification | Inkling 2 Small | Inkling-Small |
Release date | November 5, 2026 | July 30, 2026 |
Total parameters | 594B | 276B |
Active parameters | 17B | 12B |
Experts | 8 of 384 per token | Not stated in the launch post |
Context window | Up to 1M tokens | Up to 1M tokens |
Input modalities | Text and image | Text and image |
Weights | Closed for now, Apache 2.0 release announced | Full weights published at launch |
........
··········
HOW THINKING MACHINES TRAINED INKLING 2 SMALL.
The training report is unusually specific for a model of this class. Pretraining, which the lab calls phase H1, covered 8.4 trillion tokens in 68 hours on NVIDIA GB300 NVL72 racks and ran natively in the NVFP4 four-bit floating-point format. A second phase, H2, added 3.8 trillion tokens of newer and higher-quality data at 512K context. Phase H3 used about 320 billion tokens of reasoning and agent trajectories at 1M context, and phase H4 added roughly 34 billion tokens of agent trajectories.
Post-training then moved to a single joint reinforcement-learning run of 1,075 steps, with more than 17 million rollouts over nine days on 2,048 GPUs. The company puts the compute of that run above 4 × 10²³ FLOPs and the tokens processed above 3 trillion. The model card lists MXFP8 and NVFP4 as the numeric formats. It does not state a training data cutoff date, saying only that the model's knowledge is limited to what was available at its cutoff.
The lab also published a fine-tuning result that is relevant to Tinker customers. On a financial-document summarization task, Inkling 2 Small reached the final score of Inkling-Small after just 75 training steps. Thinking Machines presents this as evidence that the new model adapts faster, but it published the result for one task, so it should be read as a single data point rather than a general rule about fine-tuning efficiency.
··········
INKLING 2 SMALL BENCHMARKS AGAINST DEEPSEEK, KIMI AND GLM.
All the scores below come from Thinking Machines' own announcement and model card, and some of them are scores that the company says it took from external sources such as Artificial Analysis, Vals, Scale AI and Arena. Competitor figures are the ones Thinking Machines chose to print in the same rows. No independent evaluation of Inkling 2 Small had been published at the time of writing.
........
Benchmark | Inkling 2 Small | Best competitor cited by the lab |
AA Intelligence Index v5.0 | 30.6 | Not given in the same row |
FrontierCode v2.0 | 25.4% | DeepSeek-V4.1-Flash-1017: 28.5% |
SWEBench Pro v2 (Hard) | 86.3% | Kimi K3: 88.2% |
Vibe Code Bench v1.1 | 81.6% | DeepSeek-V4.1-Flash-1017: 87.7% |
DeepSWE v1.1 | 72.1% | DeepSeek-V4.1-Flash-1017: 77.0% |
MMMU Pro (Standard 10) | 81.6% | Claude Haiku 5.5: 79.1% |
ScreenSpot Pro | 83.8% (91.5% with Python) | Highest in the lab's table |
HLE-Diamond | 28.7% | Kimi K3: 26.9% |
AA-LCR v1.1 | 83.3% | Kimi K3: 88.7% |
GDPVal-AA v2.1 | 1630 | DeepSeek-V4.1-Flash-1017: 1746 |
PostTrainBench v1.2 | 31.9% | GLM-5.3-Flash: 36.8% |
........
The pattern is consistent. Inkling 2 Small leads in multimodal and interface work: it posts the highest ScreenSpot Pro score in the table, the top open-weights DesignArena Website score at 1340, a DesignArena Web Apps score of 1296, and 90.2% on BrowseComp with context management. It also edges Kimi K3 on the HLE-Diamond reasoning set and Claude Haiku 5.5 on MMMU Pro. In agentic coding and knowledge-work tasks it usually sits a few points behind DeepSeek-V4.1-Flash-1017 or Kimi K3, and on SWE-bench Verified the model card gives 82.2% against a top score of 95.4% in the same table.
A direct comparison with the first-generation model on the headline index is not possible from the published figures. Inkling-Small was reported at 40.0 on version 4.1 of the Artificial Analysis index, while Inkling 2 Small is reported at 30.6 on version 5.0, and the two versions use different test sets and scales. The lab describes the new model as "more capable across the board" than the first generation, which is a vendor claim that the version change prevents readers from checking against the index alone.
··········
INKLING 2 SMALL PRICING ON TINKER AND WHAT IT COSTS TO RUN.
Tinker publishes two sets of prices for the model, both in USD per million tokens. Serverless inference, which is in beta, costs $0.30 for input, $0.03 for cached input and $1.00 for output on a 256K context. The training and sampling price list, dated October 19, 2026 on the pricing page, charges the same $0.30 prefill, $0.03 cached prefill and $1.00 sampling rates, plus $1.50 per million training tokens at 64K context and $3.00 at 256K. Tinker prices MoE models by active parameters rather than total size, which explains why a 594B model is listed below several smaller dense and MoE models. Checkpoint storage costs $0.10 per GB per month.
........
Model on Tinker | Prefill (input) | Cached prefill | Sample (output) | Train, 64K | Train, 256K |
Inkling 2 Small | $0.30 | $0.03 | $1.00 | $1.50 | $3.00 |
Inkling-Small | $0.58 | $0.116 | $1.44 | $1.73 | $3.47 |
DeepSeek-V4.1-Flash | $0.60 | $0.12 | $3.60 | n/a | $4.98 |
GLM-5.3-Flash | $0.84 | $0.168 | $2.11 | n/a | $5.58 |
Kimi-K3.1-Pro | $6.00 | $1.20 | $25.00 | n/a | $32.00 |
........
Prices as listed on the official Tinker pricing page on November 7, 2026. The other models' rows are Tinker's own rates for hosting them and do not reflect the prices of DeepSeek, Zhipu or Moonshot on their own APIs. On the serverless inference list, Inkling-Small costs $0.30 input, $0.06 cached input and $1.20 output, so the gap between the two generations is narrower there than on the sampling list.
To give the rates a concrete shape, consider a coding-agent session of 40 turns, each sending 60,000 input tokens of which 85% hit the cache, and each producing 3,000 output tokens. The session totals 2.4 million input tokens, split into 0.36 million uncached and 2.04 million cached tokens, plus 0.12 million output tokens. At Tinker's sampling rates, the session costs about $0.29 on Inkling 2 Small, $0.62 on Inkling-Small, $0.89 on DeepSeek-V4.1-Flash and $0.90 on GLM-5.3-Flash. Using the serverless rate for Inkling-Small instead, the older model comes to about $0.37, so Inkling 2 Small is roughly 23% cheaper for this workload on that list. A team running 1,000 such sessions a month would spend about $289 on Inkling 2 Small at these rates.
Fine-tuning follows the same direction. A 20-million-token training job costs $30 at 64K context and $60 at 256K on Inkling 2 Small, against $34.60 and $69.40 on Inkling-Small. Readers comparing these numbers with first-party APIs can find DeepSeek's own pricing in our coverage of DeepSeek V4.1 Flash and its lower prices, and Moonshot's in our report on Kimi K3.1 Pro and its 1M-token context.
··········
WHO CAN USE INKLING 2 SMALL AND WHEN THE WEIGHTS ARRIVE.
Access today is limited to Tinker. Developers can try the model in the inference-only Tinker Playground, call it through serverless inference, sample from it, and fine-tune it with training jobs billed per token. The model card says the model is distributed under the Tinker Terms of Use for now. Thinking Machines states that the model "will be available on additional platforms in the coming weeks" once safety testing is complete, without naming those platforms.
The open-weights plan is a stated intention with a short deadline. The announcement says the lab intends to publish the full weights under Apache 2.0 "next week", with downloadable checkpoints, and elsewhere describes the weights as "pending further safety evaluations by us and our partners." Until the files are public, Inkling 2 Small is a closed-weights model, and the license terms should be checked again on the day of release. Apache 2.0 would make it more permissive than many recent large open-weight releases, which ship under custom or modified licenses. Earlier context on the lab's funding and its first Inkling models is in our report on Thinking Machines Lab's funding talks and the Inkling family.
For teams evaluating the model, the practical sequence is to test it on Tinker at the 64K or 256K tier, compare cost per completed task against their current open-weight model rather than per-token price alone, and plan any self-hosting only after the weights and their license are actually published. A 594B MoE with 17B active parameters still needs enough accelerator memory to hold every expert, so self-hosting will require multi-GPU servers even though the per-token compute is close to that of a 17B dense model.
··········
INKLING 2 SMALL SAFETY RESULTS AND RELEASE CONDITIONS.
Thinking Machines ties the weights release to safety work, and it published several safety scores. On its internal misbehavior measure, Inkling 2 Small scores 14.9 against 66.8 for the original Inkling, while the closed-weights GPT-6.1 Sol scores 0.8 on the same test. On the FORTRESS benchmark, where lower is better, its average risk score is 16.2, below the 26.7 of GLM-5.3-Flash and the 27.2 of DeepSeek-V4.1-Flash. The trade-off shows in the overrefusal score of 20.4, compared with 3.6 for GLM-5.3-Flash, which means the model declines more benign requests than that competitor.
The model card states that loss-of-control testing found the model "materially below frontier capabilities", and that evaluations covered sycophancy, manipulation, psychological harm, chemical, biological, radiological and nuclear risks, and cyber risks. It also lists the usual limitations of foundation models, including hallucination, possible bias, uneven performance across languages and domains, and weaker performance in long multi-turn conversations, and it advises against medical, legal or safety-critical deployment without further fine-tuning and validation.
··········
INKLING 2 SMALL: CONFIRMED FACTS AND OPEN QUESTIONS.
Several points are established by the company's own documentation: the 594B-total and 17B-active MoE design with 384 experts, the 1M-token context, text and image input, the training schedule on GB300 hardware in NVFP4, availability on Tinker from November 5, and Tinker prices of $0.30 input, $0.03 cached input and $1.00 output per million tokens. The benchmark results are vendor-reported, and the comparison rows were selected by Thinking Machines.
Three questions remain open. The first is whether the Apache 2.0 weights appear on the announced schedule and under the announced license. The second is how the model performs in independent evaluations, since the published table shows it ahead in browsing, vision and web design but behind DeepSeek-V4.1-Flash and Kimi K3 in most agentic coding rows. The third is which additional platforms will host it and at what prices. For now, Inkling 2 Small is a low-cost, fine-tunable model on a single platform, with a credible but still pending path to open weights.
··········
FOLLOW US FOR MORE.
·····
DATA STUDIOS
·····
datastudios.org
