News · · 8 min read

Inkling 2 Small: A 594B Model That Ships on an API Before Its Weights, and What an Independent Score of 31 Means

In this post
  1. What changed from the first generation
  2. Position among open models of similar active size
  3. Fewer hallucinations, and the price paid
  4. Where the weights will fit
  5. Developer reactions
  6. What this changes for developers in Korea
  7. Limitations
  8. Sources

Thinking Machines Lab released Inkling 2 Small on November 5, 2026. It is a mixture-of-experts (MoE) model with 594B total and 17B active parameters, and it accepts image input. Artificial Analysis, an independent evaluator, measured it at 31 on its Intelligence Index v5.0.1, up 18 points from the first-generation Inkling Small (13) released in July. The release order is unusual, though. For now the model runs only on the company's Tinker platform with closed weights. Thinking Machines says it will publish the weights under Apache 2.0 "next week," once external safety testing is complete.

That leaves two questions a developer can answer today: where the model sits among open models with a similar active parameter count, and whether the weights will fit on hardware you can actually own. This post covers what changed from the first generation, the independent index set against active parameters, how the hallucination rate fell and what it cost, a memory calculation for the weights, and developer reactions.

What changed from the first generation

Inkling 2 Small is Thinking Machines' second model generation. The company says it changed every stage from pre-training to reinforcement learning. The first generation ran separate reasoning and agentic RL runs and merged them with distillation; this one used a single joint RL run (1,075 steps over 9 days). The company also states that the data in its last two training phases consisted mostly of reasoning and agent trajectories (agent trajectories only in the final phase) sampled from Kimi K3 and its own earlier models, filtered for correctness and alignment.

Artificial Analysis measured both generations on the same index. Parameter counts are Thinking Machines' official figures (Artificial Analysis lists Inkling Small at 266B):

MetricInkling SmallInkling 2 Small
Total / active parameters276B / 12B594B / 17B
Intelligence Index v5.0.11331
AutomationBench-AA5%46%
AA-Omniscience hallucination rate61%27%
Output tokens per index task~37,000~150,000

Agentic work improved the most. AutomationBench-AA went from 5% to 46%, and AA-Briefcase, a long-horizon business benchmark that has models produce spreadsheets, presentations and memos, went from 912 to 1,527 Elo. According to the company, pre-training ran natively in NVFP4, the 4-bit format introduced with NVIDIA Blackwell, and the first phase (8.4 trillion tokens) took 68 hours on GB300 NVL72 racks.

Position among open models of similar active size

An MoE model computes only some experts per token, so latency and serving cost track the active parameter count closely. We pulled index scores and parameter counts from Artificial Analysis's model pages and put them on one axis alongside open-weight models we have covered before.

Figure 1. Artificial Analysis Intelligence Index v5.0.1 scores with active / total parameters. All values from Artificial Analysis model pages, checked November 6, 2026. Labels in parentheses are the reasoning settings Artificial Analysis lists.

Inkling 2 Small is not on top. DeepSeek V4.1 Flash 1017, with a similar 16B active, scores 33 at its Max setting. Inkling 2 Small does beat Mistral Large 4 (26) and Muse Spark Open (25), with 2.5 to 3 times as many active parameters. That position is why Artificial Analysis says the model, once its weights ship, would be the highest-scoring open-weights model developed outside China.

One caveat applies. Artificial Analysis currently classifies Inkling 2 Small as proprietary. Until the weights actually appear, it is not on equal footing with the other models in Figure 1.

Fewer hallucinations, and the price paid

The drop from 61% to 27% comes from the model answering less, not knowing more. Artificial Analysis reports that accuracy on AA-Omniscience rose only from 32% to 34%, while the share of questions the model attempted fell from 77% to 53%. The model learned to decline questions it is unsure about. The company says it used uncertainty-sensitive rewards in RL, and its limitations section notes a higher over-refusal score.

The second price is time and tokens. Artificial Analysis measured 21.2 minutes per index task, against a 7.0-minute median for its comparison group. It produced about 150,000 output tokens per task, four times the first generation, and took 36.24 seconds to the first answer token. Token prices are low at $0.30 per million input and $1.00 per million output, but the long outputs push the cost per task to $1.18, almost six times Inkling Small's $0.20.

Taken together, Inkling 2 Small looks built for long agentic jobs that can run for tens of minutes rather than for snappy chat. If users wait on screen for its answers, measure that wait first.

Where the weights will fit

Since the weights are promised, we calculated whether 594B parameters fit on real hardware. Per NVIDIA's documentation, NVFP4 stores each value in 4 bits and adds one 8-bit scale per 16 values, an average of 4.5 bits per parameter. Every parameter has to sit in memory, not just the active ones, so the calculation uses total parameters.

ModelTotal parametersNVFP4 weights (our calculation)Fits in 2x DGX Spark (256GB)?
Inkling 2 Small594B~334GBNo
DeepSeek V4.1 Flash 1017552B~311GBNo
GLM-5.3-Flash320B180GBYes
Inkling Small276B~155GBYes

The formula is total parameters × 4.5 bits ÷ 8, with 1GB as one billion bytes. This is a lower bound for weights alone; it leaves out KV cache and runtime memory. An NVIDIA DGX Spark has 128GB of unified memory in its larger configuration, so even two of them cannot hold Inkling 2 Small at 4 bits. A Hacker News commenter suggested a Q2–Q3 quantization could run on two Sparks. By our calculation, 3 bits brings the weights to about 223GB, which fits, but quantization formats above 3 bits push back toward 256GB, and quality loss comes with it.

By the same calculation, the first-generation Inkling Small comes to about 155GB and fits on two Sparks. The second generation has more than twice the total parameters, so teams that sized hardware for the first generation should redo the math.

Developer reactions

The Hacker News thread for the launch (in the high 180s for points, with 65 comments, as of November 6) mixed enthusiasm and skepticism.

On the positive side, commenters focused on having a capable open model from a US lab. User Jackobrien wrote:

"The best open models today have been "good enough" to do most knowledge work for a few months. This might be the first US one good enough too." (Hacker News)

User a11r liked that the company's own benchmark table does not show it winning every row:

"Very refreshing to see benchmarks where the reporting team is not the clear winner." (Hacker News)

The skeptics compared it with DeepSeek V4.1 Flash. User lifeisloving pointed to price:

"If you cant beat Deepseek on price im not using it." (Hacker News)

User alexandra_au questioned performance per parameter:

"It doesn't beat DS v4.1, and they have so much more compute capacity than any Chinese company, yet still cannot even beat MiniMax M3.1 Flash (220B) despite having 2.7x more parameters (596B)." (Hacker News)

The figures in that comment do not match the official ones. Inkling 2 Small has 594B parameters, and Artificial Analysis lists MiniMax-M3.1-Flash at 428B. On the index, Inkling 2 Small (31) is also ahead of MiniMax-M3.1-Flash (29). The point that it trails DeepSeek V4.1 Flash 1017, however, matches Figure 1.

What this changes for developers in Korea

Little changes right away. At launch the model is offered on the Tinker platform and Tinker Playground. The model card lists its languages only as "English, general multilingual support," and neither the company nor Artificial Analysis has published a Korean evaluation. Anyone planning a Korean-language product will have to test it on their own Korean tasks.

If the weights do ship under Apache 2.0, there will be one more option. On score alone, DeepSeek V4.1 Flash 1017 still leads in the same active-parameter range, so the reason to pick Inkling 2 Small seems more likely to be conditions than raw capability: organizations that need to weigh where a model comes from, or teams that want to fine-tune on Tinker and then download the weights to serve in-house. The company names fine-tunability as a goal of this model, but the only comparison so far is its own experiments.

The long response times narrow the fit. Rather than features that must answer in a few seconds, such as payments or customer support, overnight batch agent work like document analysis, code migration and report writing looks like the better place to try it first.

Limitations

ItemDetail
No direct runsWe did not call the model ourselves. Comparisons compile Artificial Analysis's published values
Weights not yet releasedAs of November 6 the weights are not out, and Artificial Analysis lists the date as unconfirmed, and the company's Inkling overview page says "in the coming weeks." We cannot confirm the released weights will match the Tinker model
Different reasoning settingsArtificial Analysis measured the models in Figure 1 at different settings (Max, High, Xhigh)
Memory calculationLower bound for weights only. Excludes KV cache, runtime memory and the actual release file format
Parameter discrepancyThinking Machines gives Inkling Small 276B total parameters; Artificial Analysis lists 266B. We use the official figure
Measurement dateArtificial Analysis figures such as speed and task time change over time. Values checked November 6, 2026
Korean performanceNo Korean evaluation exists from the company or from independent testers

What generalizes is the direction, "near the top of its active-parameter class but not first, and slow," not the individual numbers.

Sources

1 / 1