AI News · 4 min read
Inkling 2 Small: 594B MoE, open weights next week
Thinking Machines Lab shipped Inkling 2 Small on Nov 5, 2026 — 594B MoE / 17B active, AA Index 31, live on Tinker now; Apache 2.0 weights intended next week after safety testing.
What Thinking Machines shipped on November 5, 2026
Thinking Machines Lab published Introducing Inkling 2 Small ↗ on November 5, 2026, with a matching model card ↗ and Tinker models & pricing ↗.
The short version:
594B total MoE / 17B active parameters; image inputs and variable effort; no audio.
Artificial Analysis Intelligence Index score 31 (vendor / AA — attribute; official post says the predecessor Inkling Small scored 13).
Live now on Tinker for inference and fine-tuning; weights are closed at launch.
They intend to release full weights under Apache 2.0 next week after external safety testing and red-teaming — that is a conditional promise, not a download link today.
Tinker pricing from the docs table (per million tokens): Prefill $0.30 / $0.03 cached, Sample $1.00, Train $1.50 at 64K; the 256K variant trains at $3.00.
Notable vendor-reported benches from the announcement (attribute to Thinking Machines / named boards — not our lab):
SWEBench Pro v2 (Hard): 86.3%
EnterpriseBench-CoreCraft: 71.8%
DesignArena Website: 1340 (November 5 leaderboard snapshot per their post)
Official limitations they state: verbosity, imperfect tool selection when many skills are present, weaker chat organization / presentation, overreaction on some emotional-support and safety queries, and no audio.
Context windows on Tinker IDs in the pricing table are 64K and 256K variants. Their eval write-up mentions a 1M-token maximum trajectory limit for benchmarking with compaction — do not treat that as “1M context shipping in the default Tinker SKU” without checking the Tinker ID you actually call.
What people are saying on X
Official launch post:
“Introducing Inkling 2 Small. Scoring 31 on the AA Intelligence Index, it's a big jump over its predecessor.
Inkling 2 Small is available for fine-tuning and inference on Tinker today, with full weight release planned soon.”
— @thinkymachines ↗ (~5.1k likes / ~1.4M impressions at research time)
Competitive “US open-weights” framing — anticipation, not a download today:
“Overall, it's the first time an American open-weights model has felt competitive with the Chinese open-weights ecosystem to me. A win for open models.
Congratulation to the @thinkymachines team! Really impressive work :)”
Caveat worth repeating: weights are still closed. The Apache 2.0 drop is “next week” if safety testing clears. Treat open-weights celebration as forward-looking until the files are public.
Closed-today / open-next-week summary:
“Thinking Machines' Inkling 2 Small is live — 594B MoE, 17B active, 1M context, built for long agentic rollouts (median 8 min, 99th percentile 4 hrs).
Catch: weights are still closed. Apache 2.0 only next week, after safety testing. Tinker-only for now.”
On that “1M context” line: prefer the official Tinker 64K / 256K SKUs for pricing and product claims; the announcement’s training notes cite long-horizon rollouts (median ~8 minutes, p99 ~4 hours) and a 1M-token eval trajectory ceiling — attribute carefully.
Our take: who should care
Tinker fine-tuners and agent teams now: you can sample and train Inkling 2 Small today without waiting for weights. The MoE shape (17B active) plus vision + variable effort is aimed at long agentic coding and knowledge-work loops — that is the product story, not a chat-UI polish pass (they admit chat organization is weaker).
Self-host / open-weights waiters: calendar the “next week + Apache 2.0” promise, but gate procurement on the actual weight drop and license file. Until then it is Tinker-only closed weights. Do not ship a self-host runbook that assumes downloadable checkpoints today.
Distinct from Kimi’s open-weights path: we already covered Kimi K3.1 Pro open weights. Inkling 2 Small is a US-lab MoE play with a delayed open drop; Kimi’s open release is a different timeline and stack. Compare on your own harness — do not paste vendor benches into a “winner” headline.
Action this week: try the Tinker playground / fine-tune path if you need US-lab agentic coding + vision now; watch the safety-gated weight release before planning offline deploy. Price the 64K vs 256K train SKUs ($1.50 vs $3.00 train) before you burn a long RL run.
For the wider open-model landscape see best open-source LLMs 2026. For coding assistants see best AI for coding 2026. For agent stacks see best AI agent frameworks 2026.
AdSense-safe note: we are documenting what Thinking Machines published and how to evaluate the Tinker vs open-weights gate — not an upsell for any paid Tinker plan.