Inkling Small: open 12B-active model beats bigger Inkling at coding
Apache-2.0, 276B total but 12B active, 1M context — near-flagship scores at a quarter the compute. Plus: GitHub Copilot drops two Gemini models.
Copy markdown
12B active — and it out-codes the 41B-active original
Inkling Small routes each token to 6 of 256 experts, so just 12B of its 276B params fire — yet it hits 80.2% on SWE-bench Verified, above full Inkling's 77.6%, plus 64.7% on Terminal-Bench 2.1 and 89.5% GPQA Diamond. Near-flagship agentic coding at a quarter the active compute.
Apache 2.0, 1M context, multimodal in
Weights ship under Apache 2.0 with a 1M-token window and text, image, and audio input. No gated license and no usage caps — fine-tune it, embed it in a product, or self-host it without asking permission.
One box self-hosts it, or call the API cheap
The NVFP4 quant fits about 180GB (versus 600GB in BF16), so a single high-memory node runs it locally. Prefer hosted? It's on Tinker and third-party providers, with a launch promo near $0.58 in / $1.44 out per million tokens.
Build this weekend: a coding agent you fully own
80.2% SWE-bench Verified at 12B active means a genuinely capable agent you run on your own hardware — wire it into your editor or CI and keep every token in-house. One caveat: it's weak on tool-heavy financial flows (Tau-cubed Banking just 15.5%).
Elsewhere: GitHub Copilot drops Gemini 2.5 Pro and 3 Flash
As of July 31 both models are deprecated across every Copilot surface — chat, agents, inline edits, completions. Move workflows to Gemini 3.1 Pro or Gemini 3.6 Flash; the old ones still respond for now but are no longer supported.