Open Source · NOVEMBER 21, 2026
Inkling 2 Mini brings looped agentic coding to local hardware
Thinking Machines’ open-weight model fits compatible local machines, but its coding scores are not evidence of better customer-acquisition performance.
Thinking Machines Lab released Inkling 2 Mini on November 20, a 26.4-billion-parameter open-weight agentic coding model with a reported roughly 20GB total footprint.
Weights and Tinker inference/customization access are available. The lab reports 48.2% on DeepSWE v1.1; that isn't evidence of better customer acquisition.
Repeated layers, not extra words
The 24-layer mixture-of-experts model repeats its middle 12 layers three times per token: 48 layer applications, 3.4 billion active parameters and 5.4 billion weight activations per token including looping. Expert pruning reduced 384 experts to 128 per layer; training included 1 trillion distilled tokens and 2.2 trillion tokens over 2,000 steps.
Goodfire, given early weights, analyzed attention residuals and cross-loop attention across 13,233 tokens from 400 FineWeb documents. Later passes can read preceding-loop outputs; later loops generally favored the third block’s outputs over the second’s.
Separate TML-d24-varloop-research tests used one to five passes. At 50% reliability, the no-chain-of-thought time horizon rose from 16 seconds with one pass to 37 with three. Serial reasoning improved through ten loops, twice the training maximum. Thinking Machines reports shorter, legible traces without increasing agentic sabotage. Extra passes helped long-context results, not knowledge-intensive AA-Omniscience or SimpleQA Verified.
Memory and serving conditions
SGLang lists about 18GB of weights; that isn't total memory. Thinking Machines trained and deployed 20 layers in NVFP4 and four in MXFP8, demonstrating offline flight-simulator creation on a 48GB M5 Pro MacBook Pro.
SGLang supports RTX 5090 (32GB), RTX PRO 6000 Blackwell (96GB), and DGX Spark (128GB shared CPU/GPU memory). Its RTX 5090 test reports 350 tokens/second and 492 milliseconds to first token with random 8K-token inputs, 1K-token outputs, and simulated speculative acceptance of 3.5 tokens/step.
InferenceX AgentX coding-trace replay reports p90 interactivity of 168 tokens/second on one RTX PRO 6000 versus 23.6 on one DGX Spark, with one user and the same simulated acceptance. These aren't unconditional speed promises.
SGLang lists text/image input, text output, tool calling and streaming through Chat Completions API, Responses API and Anthropic Messages API.
Coding scores aren't conversion results
Thinking Machines’ self-reported results include:
- SWE-bench Pro v2 Hard: 77.9%; DeepSWE: 48.2%, both using mini-swe-agent v2.4.6. Qwen3.8-27B scored 73.8% and 42.2%, respectively; evaluation setups vary by model.
- SWE-bench Verified: 79.2%; the release also identifies 51 tasks.
- BrowseComp: 79.0%; DeepSearchQA: 85.3%, both with compaction. MCP Atlas Public: 82.7%.
- MultiChallenge: 50.0%; ComplexConstraints v1.1: 33.4%, exposing multi-turn and many-constraint weaknesses.
- AA-Omniscience v1.1 non-hallucination: 63.4%, versus Qwen’s 71.4%.
For a technically capable small-team owner with compatible hardware, a hypothetical trial could revise a landing page or lead-capture form against explicit acceptance criteria, then test real visitor and inquiry behavior. That's a proposed evaluation method, not a reported result.
"The practical implication: test the inquiry flow, not just whether the model finishes the code."
Neither coding benchmarks nor offline demonstrations establish more leads, better-qualified inquiries or increased sales.