DEV Community

Cover image for Inkling 2 Small vs Muse Spark Open: Best US Open Model
Shamuddin
Shamuddin

Posted on Originally published at aitecharchive.com

Inkling 2 Small vs Muse Spark Open: Best US Open Model

Inkling 2 Small is the best US open-weight model for data analysis that Plotly has tested, and it beats Meta's Muse Spark Open by one question. Thinking Machines Lab published the weights on November 12, 2026 under Apache 2.0, and Plotly's 43-question benchmark shows it passing 39 (91%) against 38 for Muse Spark Open, at $1.81 per full run against $5.16. If you want an open-weight model from a US lab, Inkling 2 Small is the one to pick. If raw accuracy matters more than origin, Kimi K3.1 Pro still scores higher.

TL;DR

How does Inkling 2 Small compare with other open-weight models?

Plotly's benchmark is a data-analysis exam built from real-world tasks. The table below uses its published figures; the price conversion uses Rs 94.05 per dollar.

Model Score (of 43) Cost per run Median time per question Open weights
Kimi K3.1 Pro 42 not stated here about 2.5 minutes Yes
MiniMax M3.1 Flash 41 not stated here about 3 minutes Yes
Inkling 2 Small 39 $1.81 (Rs 170) 13 seconds Yes
Muse Spark Open 38 $5.16 (Rs 485) 48 seconds Yes
Inkling-Small 30 $1.93 (Rs 182) 25 seconds Yes
Particle-1 40 $0.35 (Rs 33) 16 seconds No

Source for all rows: Plotly. Two readings matter. First, within open weights, Plotly says no model is both cheaper and more accurate than Inkling 2 Small, which puts it on the cost-accuracy frontier. Second, across all models the closed Particle-1 scored 40 for $0.35 per run (Plotly), so Inkling 2 Small is not the overall bargain. Against its predecessor the gain is nine questions for about the same money, and the run takes roughly half the time.

For a fuller look at the accuracy leader, see our coverage of the Kimi K3.1 Pro open-weights release.

Where does Inkling 2 Small do well, and where does it fail?

Per the Plotly results, it scored 5/5 on impossible requests, 3/3 on data-quality problems and 5/5 on statistics traps. In practice that means it declined questions the data could not answer, caught all three planted data issues, and avoided every statistical trap.

It lost four questions. One was an analytics question where it wrote its final table call as text instead of running it. The other three were hidden-needle questions, where an answer sits inside a large body of material; it found 2 of 5. Muse Spark Open also found 2, while Kimi K3.1 Pro and MiniMax M3.1 Flash found all five. If your workload is retrieval from long documents, this is the limitation to test first.

Efficiency helps explain the low bill. Plotly measured about 5,700 output tokens per question, against about 7,800 for Muse Spark Open and 32,000 for MiniMax M3.1 Flash, and about 96% of input tokens came from the provider's cache (see the same report).

Thinking Machines is open about the other weaknesses. In its announcement, it says the model is less tuned for chat, with weaker content organisation, overuse of slashes, and overreaction to mental-health and safety-related queries.

What does Inkling 2 Small cost?

Hosted access through OpenRouter, with Together and Cloudflare serving, is priced as follows, according to Plotly. Rupee figures use the Rs 94.05 rate as of November 13, 2026.

Item USD INR
Input, per million tokens $0.30 about Rs 28
Output, per million tokens $1.00 about Rs 94
Context window 1M tokens not applicable

For a price comparison against coding-focused rivals, see Ripple 65B versus open-weight coding models.

How do you run Inkling 2 Small yourself?

There are four routes, from least to most effort.

  1. Try it hosted. The Tinker Playground is free to try, and Tinker also handles fine-tuning and sampling.
  2. Call a managed API. OpenRouter has a paid endpoint, and the Hugging Face page lists providers including NVIDIA, Fireworks, Together, Baseten, Nebius, Modal and Cloudflare.
  3. Download the weights. The NVFP4 checkpoint and the MXFP8 checkpoint are both Apache 2.0.
  4. Serve it locally with an official recipe: vLLM or SGLang. Unsloth and TokenSpeed are also supported, and Miles can post-train it.

On hardware, the model card calls for a GPU cluster with at least 500GB of aggregated VRAM for the main checkpoint, with two NVIDIA B300 cards as an example. Plotly says the 4-bit version needs about 356GB, which fits a single 512GB Mac Studio. The MXFP8 build runs on one server node.

Reported serving figures, from measurements by Thinking Machines and Inferact: the model has 48 layers and 384 experts per layer with 8 active per token. A single user gets about 630 tokens per second on four GB200 GPUs with speculative decoding. On an agentic workload across 8 GPUs, throughput falls from about 1,070 tokens per second per GPU at 32K context to about 450 at 256K, as concurrent users drop from roughly 84 to 37 (AI Aulis, citing the vLLM blog measurements).

If MIT-licensed self-hosting is your priority, compare it with LongCat 2.5.

How does it fare against closed frontier models?

On GPU kernel optimisation, the company's own benchmark shows a gap. On a DeepSeek-V3.2 sparse-attention kernel, Inkling 2 Small reached a 6.02x speedup, against 18.80x for Claude Opus 5.5 and 10.26x for GPT-6 Astra, and these apply to that kernel alone, not end-to-end training. Its 494.27 AUC on PMPP-Hard was the second-best reported in that row. For closed-model context, see our pieces on Claude Fable 5.2 versus 5.1 and Hydrogen 1.0 versus GPT-6 Astra.

Which model should you pick?

Pick Inkling 2 Small if you need an open-weight model from a US lab, want quick answers at low cost, and your work is analysis rather than needle-in-a-haystack retrieval. Pick Kimi K3.1 Pro or MiniMax M3.1 Flash if accuracy matters most and you can wait minutes per question. Skip Muse Spark Open for this workload, since it is one point behind and costs about three times as much per run. The Artificial Analysis score of 31, against 29 for Inkling and 30 for Inkling-Small, points the same way: a real step up, not a leap.

FAQ

Q: Is Inkling 2 Small really open source?
A: The weights are open under the Apache 2.0 licence, as announced on November 12, 2026. Training data and code are not covered by that announcement, so "open weights" is the accurate term.

Q: How big is Inkling 2 Small?
A: It has 594B total parameters with 17B active per token, and handles image and text inputs. The context window reaches 1M tokens according to Plotly.

Q: Can I run Inkling 2 Small on one machine?
A: Yes, with limits. Plotly says the 4-bit build needs about 356GB, which fits a 512GB Mac Studio, and the MXFP8 build runs on a single server node. The model card lists a minimum of 500GB aggregated VRAM for the main checkpoint.

Q: How much does the API cost in rupees?
A: About Rs 28 per million input tokens and Rs 94 per million output tokens, converting the $0.30 and $1.00 prices at Rs 94.05 per dollar.

Q: Is it good for chat assistants?
A: Not ideal. Thinking Machines says it is less optimised for chat, with presentation issues and overcautious safety responses. It suits analysis and agent work better.

Top comments (0)