Every model on the bench: tested with full scores, testing in progress, or partial where the lane vanished mid-run. New models appear here first.
Liquid AI's tiny LFM 3 1.7B: IFEval-50 83.7%, GSM8K-50 84.0% with zero failures, SVG 10/50 over 1 judged test. Heavy reasoning traces burned whole budgets: 1 IFEval prompt and 4 of 5 SVG briefs unrecoverable even at 32k, counted as failures. Strong small-model reasoning signal, weak long-form emission.
Nvidia Nemotron 3.5 Lightning free lane: IFEval-50 89.8%, GSM8K-50 70.0% with zero failures, SVG 55/250 over 5 judged tests. Best IFEval on the board so far, weakest GSM8K. The clock brief scored 0: the model emitted JavaScript-generated SVG instead of static shapes, and scripts are stripped before render.
Nvidia Nemotron 3 Nano Omni 30B reasoning free lane: IFEval-50 89.8%, GSM8K-50 78.0% with zero failures, SVG 104/250 over 5 judged tests. Ties Lightning on IFEval, beats it on math and SVG. The small reasoning model is the best-balanced free lane on the board so far.
Cohere North Mini Code free lane: IFEval-50 91.8%, GSM8K-50 80.0% with zero failures, SVG 98/250 over 5 judged tests. New best IFEval on the board; the code specialist also drew the best pelican yet (34/50). Its first clock attempt was 20k chars of Python that never emitted SVG, redone cleanly at 32k.
Kilo Auto router (anonymous free tier): IFEval-50 87.8%, GSM8K-50 92.0% with zero failures, SVG 145/250 over 5 judged tests including the best clock on the board (43/50). Collection survived the anonymous promo limit and AtlasCloud 400s; 1 IFEval prompt unrecoverable, counted as failure. Router product: scores describe the route, not one model.
Beam free preview via OpenCode Go: GSM8K-50 84.0% (42/50, zero failures) through the OpenCode agent harness, so the number is not strictly comparable with the direct-prompted lane. IFEval and SVG are not run for harness models: the agent harness breaks their comparability.
Iceman contributor free preview via OpenCode Go: GSM8K-50 80.0% (40/50, zero failures) through the OpenCode agent harness, so the number is not strictly comparable with the direct-prompted lane. IFEval and SVG are not run for harness models: the agent harness breaks their comparability.
Liquid AI's small LFM 2.5: IFEval-50 87.8%, GSM8K-50 74.0% with zero failures, SVG 68/200 over 4 judged tests. The clock brief proved unrecoverable (reasoning exhausted a 32k budget twice), counted as failure. The spike's small-model efficiency contrast.
Longcat 2.5 via OpenCode Go free previews: GSM8K-50 78.0% (39/50, zero failures) through the OpenCode agent harness, so the number is not strictly comparable with the direct-prompted lane. IFEval and SVG are not run for it: the agent harness (large system prompt, unknown temperature) breaks comparability, and SVG probes entered a tool loop with no output in 240 s.
Recursive's Flint v1.0 is the spike's calibration anchor: its 93.3% IFEval-50 lands within 2 points of the vendor-published full IFEval figure (95.1%), which is what a correct harness should do on a 50-prompt sample. GSM8K-50 88.1%, SVG total 144/250 with the clock (37/50) as its best test.
Dots Studio's Dots 3 note preview carries the best GSM8K-50 in the spike at 92.0% with zero failures. No published IFEval or GSM8K reference found, so these are novel numbers. SVG total 126/250, split between the run's best clock (43/50) and its weakest monalisa (12/50).
NVIDIA's 550B-class Ultra is the spike's reference model: the established name tested alongside the lesser-known ones. SVG total 150/250 leads the board. GSM8K-50 78.0% zero-shot against 88.1% vendor eight-shot chain-of-thought on the base model, lower as expected for zero-shot.
Stealth free model that died mid-run: OpenRouter started returning 404 (no provider serves it anymore) four prompts in. 6 IFEval prompts banked at 83.3% before the lane vanished. Later revealed as an early Muse Ember.