Inkling 2 Small Hits 86.3% on SWE-Bench, 13.1% on Terminal-Bench 4

What happened

Two US model releases inside a day targeted the same developer workload: agentic coding at sub-dollar token prices. On 6 Nov 2026, MarkTechPost reported that Thinking Machines Lab released Inkling 2 Small, a sparse Mixture-of-Experts model with 594B total parameters and 17B active parameters per token. It scores 31 on the Artificial Analysis Intelligence Index, up from 13 for the first Inkling Small, and plans an Apache 2.0 open-weights release next week after red-teaming.

One day earlier, on 5 Nov 2026, VentureBeat covered Ricursive Intelligence's release of Ripple-65B Preview under an Apache 2.0 license. Trained from scratch on 12 trillion tokens, the 65B-parameter dense model is live on Hugging Face and OpenRouter at $0.2 per million input tokens, compared with $0.3 per million input tokens for Inkling 2 Small on Tinker.

Both labs published coding and terminal benchmark tables, and both tables show sharp splits depending on the evaluation harness. Thinking Machines reports 86.3% on SWE-Bench Pro v2 (Hard) and 85.1% on Terminal-Bench 2.1 using the Pi harness, yet only 13.1% on Terminal-Bench 4.0 against 33.8% for DeepSeek-V4.1-Flash-1017.

Ricursive reports 80.5% on Terminal-Bench 2.1, ahead of Reflection's Beam at 80.1%, Meta's Muse Spark Open at 76.4%, first-generation Inkling at 63.8% and Nvidia's Nemotron 3 Ultra at 56.4%. VentureBeat noted that Ricursive's model card and launch graphic list different rival scores across SWE-Bench Pro v2.0 Hard and DeepSWE because vendor tables mix internal runs with published numbers.

What the chart shows

The first chart compares Inkling 2 Small against the named rival in Thinking Machines Lab's benchmark table across five evaluations. On EnterpriseBench-CoreCraft, Inkling 2 Small leads its comparison model by 25.1 percentage points (71.8% vs. 46.7%), and on PerceptionBench it leads by 6.2 points (64.7% vs. 58.5%).

The direction reverses on newer terminal and software engineering suites. On DeepSWE v1.1, DeepSeek-V4.1-Flash-1017 leads Inkling 2 Small by 4.9 points (77% vs. 72.1%). On Terminal-Bench 4.0, DeepSeek scores 33.8% to Inkling 2 Small's 13.1%, a 20.7-point gap even though the two models sit within 5.5 points on Terminal-Bench 2.1 (90.6% vs. 85.1%).

The second chart ranks the Terminal-Bench 2.1 field from both launch disclosures. DeepSeek's September checkpoint leads at 90.6%, followed by Inkling 2 Small at 85.1% (21.3 points above first-generation Inkling's 63.8%) and Ricursive's 65B model at 80.5%.

What it means for founders

Active parameter count, not total weight count, now sets the unit economics of self-hosted and API coding agents. Inkling 2 Small activates 2.9% of its 594B total parameters on each token, letting Thinking Machines price input tokens at $0.3 per million while jumping 2.4x on the Artificial Analysis Intelligence Index (from 13 to 31).

At the same time, the 72-point drop between Terminal-Bench 2.1 (85.1%) and Terminal-Bench 4.0 (13.1%) proves that agentic coding scores are bound to the harness they were tuned on. A model that looks frontier inside a lab's Pi harness can trail by 20.7 points when the shell environment and task horizon change.

For application founders, Apache 2.0 releases at 17B active or 65B dense parameters make fine-tuning on proprietary repository traces cheap. The durable moat is not picking the top row of a launch-day table; it is owning the evaluation harness and failure-recovery loop for your specific vertical workflow.

What would change our mind

We treat launch-day coding tables as harness-specific upper bounds until independent evaluators rerun the weights under a locked scaffold. Two things would change our view. First, if Thinking Machines ships the Apache 2.0 weights next week and independent runs confirm its 86.3% SWE-Bench Pro v2 (Hard) score outside the Tinker API, Inkling 2 Small becomes the default open-weight base for coding agents.

Second, if third-party benchmarks show that Inkling 2 Small's 13.1% score on Terminal-Bench 4.0 closes toward its 85.1% Terminal-Bench 2.1 mark with a standard prompt and scaffold patch, then the gap was a harness configuration artifact rather than a reasoning ceiling.

Active parameters set the inference bill, but the evaluation harness still decides the leaderboard.

What we would back: Developer-tool and vertical-agent teams that own proprietary repository evaluation harnesses and fine-tune small active-parameter MoE weights.