
Inkling 2 Mini vs Inkling 2 Small: 11 Points for 394GB
- OrcaNEWOrcaCyber: Hy4 Flash Abliterated / Uncensored2026-11-18$0.07 / $0.28 per 1M tokens · 183 tok/s
- tencentNEWTencent: Hy4 Flash2026-11-16$0.02 / $0.07 per 1M tokens · 568 tok/s
- AlibabaNEWQwen: Qwen4 Max2026-11-16$1.50 / $4.50 per 1M tokens · 385 tok/s
- GoogleNEWGoogle: Gemini 4 Carbon2026-11-10$2.00 / $10.00 per 1M tokens · 87 tok/s
- AnthropicNEWAnthropic: Claude Fable 5.22026-11-09$8.00 / $40.00 per 1M tokens · 150 tok/s
- OrcaOrcaCyber: GLM 5.3 Abliterated / Uncensored2026-11-04$3.00 / $5.00 per 1M tokens · 99 tok/s
- OrcaOrcaCyber: Hy4 Abliterated / Uncensored2026-11-04$0.80 / $1.60 per 1M tokens · 134 tok/s
- OrcaOrcaCyber: Kimi K3.1 Pro Abliterated / Uncensored2026-11-04$2.00 / $5.00 per 1M tokens · 84 tok/s
- tencentTencent: Hy42026-11-04$0.83 / $2.50 per 1M tokens · 121 tok/s
- OrcaOrcaCyber: MiniMax M3.1 Flash Abliterated / Uncensored2026-10-29$0.80 / $1.60 per 1M tokens · 107 tok/s
- OrcaOrcaCyber: GLM 5.3 Flash Abliterated / Uncensored2026-10-27$0.80 / $1.60 per 1M tokens · 67 tok/s
- OrcaOrcaCyber: Qwen3.8 Flash Next Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 91 tok/s
- OrcaOrcaCyber: MiMo V2.6 Flash Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 41 tok/s
- MoonshotMoonshotAI: Kimi K3.1 Pro2026-10-2447Intelligence
- OrcaOrcaCyber: DeepSeek V4.1 Flash Abliterated / Uncensored2026-10-23$0.80 / $1.60 per 1M tokens · 214 tok/s
- minimaxMiniMax: M3.1 Flash2026-10-1929Intelligence
- OrcaOrca: OrcaCyber Ultra 1.02026-10-17$5.00 / $10.00 per 1M tokens · 111 tok/s
- deepseekDeepSeek: DeepSeek V4.1 Flash 10172026-10-1733Intelligence
- DeepSeekDeepSeek: DeepSeek V4.1 Flash 09102026-10-17$0.15 / $0.60 per 1M tokens · 160 tok/s
- DeepSeekDeepSeek: DeepSeek V4.1 Flash2026-10-1740Intelligence
Two checkpoints, one family, and a choice that comes down to how much disk and memory you are willing to spend on eleven Index points. Inkling 2 Mini is the small one: 26.4 billion total parameters, 3.4 billion active, a single roughly 20GB file that Thinking Machines Lab published on 20 November 2026 under Apache 2.0. Inkling 2 Small is the big one: 594 billion total, 17 billion active, 393.5GB across 36 files, weights opened on 12 November after a week of closed availability that began with its 5 November launch. On the Artificial Analysis Intelligence Index v5.0.1, which measures both independently, they score 20.1 and 31. That is the whole trade in one line, and it understates how differently the two behave.
What follows separates the two kinds of evidence, because the difference decides most arguments about this family. Index score, cost per index task, speed, context and licence are Artificial Analysis figures and vendor metadata, read on 21 November 2026. Every benchmark percentage is Thinking Machines' own evaluation table from the relevant launch post, on a checkpoint the company trained, and is marked vendor-reported each time it lands. Neither table has been reproduced outside the lab.
What each checkpoint actually is
Start with the artefact, because the artefacts are not the same kind of thing. Inkling 2 Small is a flat mixture-of-experts model: 594B total, 17B active per token, weights published eight days after launch as 24 safetensors shards plus an MXFP8 variant that is larger still. Inkling 2 Mini is a pruned, looped descendant — the same mainline recipe with two-thirds of the experts deleted and a loop added to the middle layers. It ships as one file at about 20GB because it was trained end to end in 4-bit and is deployed in the precision it was trained in, 20 layers in NVFP4 and 4 in MXFP8, so there is no conversion loss to pay for.
• Total parameters — Inkling 2 Mini 26.4B vs Inkling 2 Small 594B.
• Active parameters — 3.4B vs 17B, with Mini reaching 5.4B of effective weight activations per token once its three loop passes are counted.
• Download — about 20GB in one file vs 393.5GB across 36 files, or 622.8GB for the MXFP8 variant.
• Index — 20.1 vs 31, both on Intelligence Index v5.0.1.
• Cost per index task — $0.59 vs $1.18.
• Output speed and first token — 280 tokens per second at 9.72 seconds vs 118 tokens per second at 39.68 seconds.
• Context — 1M tokens on both model cards, and 1M recorded by Artificial Analysis for both. Inkling 2 Small's served tier on Tinker is narrower than its card, and Mini inherits that same gap; treat the million as a capability, not an operational promise.
• Licence — Apache 2.0 on both. Small's card declares it without shipping a LICENSE file; Mini's card declares it in metadata with the same absence. Worth knowing if your legal review wants paperwork in the tree.

Eleven points, and where they come from
Reading the two vendor tables side by side is the only way to see what the extra compute buys, because the index score alone hides it. Both models were evaluated by the same lab with overlapping harnesses, so the rows are roughly comparable to each other — and only to each other.
• DeepSWE v1.1 — Inkling 2 Small 72.1% vs Inkling 2 Mini 48.2%. The widest gap in the coding block.
• SWE-bench Pro v2, Hard — 86.3% vs 77.9%.
• SWE-bench Verified — 79.2% for Mini, bash only. Small's own launch table does not print the same row, so this is the one coding benchmark the two posts cannot be lined up on.
• Vibe Code Bench v1.1 — 81.6% vs 77.9%.
• FrontierCode v2.0 — 25.4% vs 15.1%. A ten-point gap on the benchmark where both are weakest.
• PMPP-Hard, the kernel-optimisation benchmark — 494.27 AUC for Small against 455.82 for Mini, which is closer than the parameter counts suggest and is the row the launch post points at when it says Mini "approaches" its larger sibling.
• BrowseComp — 90.2% vs 79.0%; DeepSearchQA 90.9% vs 85.3%; MCP Atlas 85.0% vs 82.7%. Search and tool use narrow the gap to single digits.
• CritPt v2, full — 55.7% vs 31.4%. Science reasoning is where the extra stored knowledge shows up hardest.
• AA-Omniscience — 16.1 with a 73.3% non-hallucination rate vs −8.6 with 63.4%.
That last row is the one to sit with. A negative omniscience score means a model is worse than declining to answer, and the eleven-point index gap is mostly paid for there: the larger model knows things, the smaller one works things out. Thinking Machines says this itself, in different words, when it explains that looping adds serial computation but not stored facts. The pruning cut two-thirds of the experts; knowledge lives in the experts that are left.
Same recipe, smaller pan
The family resemblance is not marketing. Both models went through SFT with a similar data mix and an RL phase totalling 2.2T tokens over 2,000 steps, with on-policy distillation used for alignment in both. The post-training pipeline is the same pipeline; what differs is what arrived at it.
Mini's route there is worth knowing because it explains several of the benchmark rows. It began at sparsity 48, routing each token to 8 of 384 experts. At one point in midtraining, Thinking Machines kept the 128 most-used experts per layer — a two-thirds cut — and healed the loss on 1T tokens distilled from a larger internal Inkling model, training the survivors to take over the deleted experts' work. Then it added looping: four blocks of six layers, with the middle two blocks run three times and a learned gate that merges the previous pass's message into the next, starting at zero gain so training begins as an ordinary model.
Two judgement calls in that design are instructive. The company looped the middle twelve layers rather than the last twelve after testing both across 28 evaluations and finding the middle better on 27, with a compute-equivalent gain of 1.51–1.91× against 1.24–1.36×. And it kept the first and last blocks single-pass, on the theory that they handle token-to-representation translation while the middle carries abstraction. That is why Mini gets 48 layer applications per token out of 24 stored layers without adding a parameter.
The cost of that trick is operational. A flat MoE is something every inference stack has served for two years. A looped model has had SGLang and vLLM support since roughly the day it shipped, and the model card says plainly that looped-model runtime support is new and that speed varies. If you are choosing between these two for a production path this month, the boring one is the one with the mature kernel.
Hardware is the real divider
The two checkpoints do not compete for the same machine, and that decides more deployments than the score does. Mini is specified to run on a single NVIDIA DGX Spark, any Apple silicon device with 32GB or more, or a GPU with 32GB of VRAM; the company's own demo has it building a playable flight simulator offline on a MacBook Pro in airplane mode, and a town simulation with twenty agents on one laptop. A Stanford efficiency study run on early access put it at 0.703 intelligence per watt on a DGX Spark at MXFP8 and low reasoning effort — measured by researchers outside the lab, on one device, which is the strongest independent signal in the release.
Small needs a node. 393.5GB of weights does not fit anywhere a person owns, which is why the natural way to consume it is a hosted endpoint — Tinker, in the vendor's case — and why the interesting comparison for most readers is not "which file do I download" but "which of these can I run at all". Mini answers that question with a laptop. Small answers it with a budget.
That asymmetry is where routing earns its place, and it is worth being concrete rather than gesturing. Neither model is on OrcaRouter, so there is nothing to route to either of them today. What is on OrcaRouter is the hosted side of the same decision: the models you would call for the workloads the 594B checkpoint is actually for, on one key, at the provider's list price with no markup added. The routing DSL is the piece that matters here — it lets you split by task so the expensive call only happens on the requests that need it, with automatic failover when a provider goes down, while the cheap local model handles everything else. Running Mini locally and calling out only for the hard tail is a smaller bill than running either model for everything, and it is the arrangement this release actually makes possible.
The decision, in three cases
If you are building an agent that lives on a developer's machine or inside a single offline box, Inkling 2 Mini is the answer and there is no real contest. A 20GB download with 280 tokens per second, Apache 2.0, and a coding table that beats every open-weights model of its size on the vendor's own rows. Sandbox it — the card warns about destructive file actions and loops while debugging — and it will do repository-scale agentic work on hardware you already own.
If your workload is retrieval over a large corpus of documents you did not write, Small's advantage is mostly in the rows that involve recalling what a document said. CritPt at 55.7% against 31.4%, omniscience at 16.1 against −8.6, and a 73.3% non-hallucination rate against 63.4% are not small differences in a compliance or finance setting, where a confident wrong answer costs more than a slow right one. You will pay for a node or an endpoint to get them.
If you are choosing for a mixed workload, the split is the point. Mini's weakness is knowledge, and knowledge is the one thing you can supply from outside — a retrieval layer, a search tool, a bigger model on the hard tail. Its strength is long-horizon agentic persistence at a price nobody else in the 20GB class matches. Buy the small model, rent the knowledge, and treat the 594B checkpoint as the thing you escalate to rather than the thing you run.

One caveat cuts across all three cases and belongs at the end rather than buried. Both models' benchmark tables are the lab's own, and the index scores are the only independently measured capability numbers in this comparison. The 31 and the 20.1 are comparable; the 86.3% and the 77.9% are comparable to each other and to nothing else. Anyone quoting the vendor rows next to a competitor's published score is comparing two harnesses, not two models.

