8.8B Model Reaches 90% of the Frontier: Don’t Rush to Rip Up Your Big-Model Bill

TwIL-8B weights have been released: 8.8B parameters, with both the model card and weights on Hugging Face. The publisher, webAI, gives the numbers directly: LiveCodeBench 81.6, reaching 93% of the top frontier score; IFBench 74.4, reaching 92%; AIME 2026 88.3, HMMT Nov 2025 86.7. The same model card also states GPQA Diamond 70.2 and HLE 12.6, with a clear lag on knowledge benchmarks.

This set of numbers matches Andrej Karpathy’s repeated judgment this year about small models. Codez relayed five points in a tweet: intelligence has never been knowledge; remove the memorized internet and what remains is the reasoning core, and this core can be very small; the core may fit into about 1B parameters; size is no longer the bottleneck—whether an agent can call tools and how tight the feedback loop is matters more; intelligence lives in the system, not in the weights; a bare model is like a CPU without an operating system; small models can be owned by you, and no one can rate-limit them, raise their prices, or take them offline. Karpathy’s original wording is sharper: “A 1B model can be much more effective than most current LLMs. 90% of current models waste a ton of capacity and money on remembering stuff that doesn’t matter.” This relay was posted one day before TwIL-8B’s release, and the point is clear.

TwIL-8B is based on IBM Granite 4.2 8B, with full-parameter SFT, then GRPO/DAPO reinforcement learning, and finally length-aware rewards to reduce reasoning length, splitting into balanced and verbose versions. balanced is the default, with shorter reasoning and lower cost; verbose is suited to hard problems and uses more tokens. The bf16 weights are 16.38 GiB, Q4_K_M GGUF is 4.98 GiB, and the context is 131,072 tokens. The model card is clear: it can run locally, knowledge is its weak spot, and it has not been trained beyond Granite 4.2 for safety, preference, or tool calling.

Putting TwIL-8B alongside open models of the same size, code and instruction following are strengths:

• MATH-500 98.2, Qwen3.5-9B 96.6, Gemma 4 31B 87.9
• AIME 2025 88.3, Gemma 4 31B 87.1
• LiveCodeBench 81.6†, Gemma 4 31B 80.0, Qwen3.5-9B 65.6
• IFBench 74.4†, Qwen3.5-9B 64.5, Gemma 4 31B 76.0
• GPQA Diamond 70.2, Qwen3.5-9B 81.7, Gemma 4 31B 84.3
• HLE 12.6, Qwen3.5-9B 14.9, Gemma 4 31B 19.5

† indicates measured on the verbose version; balanced has not been run yet. The model card also notes that competitor scores come from public sources and were not rerun. AIME and HMMT each have 30 to 33 questions, about 3 points per question, so a few points of difference fall within noise. The model card uses the highest reasoning settings of GPT-5.5, Claude Opus 4.8, and Grok 4.3 as references; TwIL-8B is closest to the frontier on LiveCodeBench and IFBench, and has the largest gap on knowledge.

The comments don’t speak with one voice. Someone asked how much accuracy drops with Q4_K_M; David Stout replied, “very little with our AIR method. better than unsloth.” Someone else directly asked “Benchmaxxing?” and the publisher didn’t elaborate. The weights are open, under the webAI Non-Commercial License ver. 1.0; commercial use depends on the license.

For ordinary users, parameter count is only one dimension. After Q4 quantization it is under 5 GiB, so the model can fit on consumer hardware, inference runs locally, data never leaves the machine, and API bills and rate limits are no one else’s business. For developers, the Hugging Face page gives loading methods for Transformers, llama.cpp, vLLM, and Ollama. With llama.cpp you can run:

llama-cli -hf webAI-Official/TwIL-8B --hf-file TwIL-8B-Q4_K_M-balanced.gguf -cnv -c 40960 -n 32768

Give the generation budget enough room; by default the model first outputs a <think> block. A short budget will truncate reasoning and scores will drop sharply. In the model card, balanced averages about 19.5k output tokens on AIME 2026 and about 32.9k on HMMT Feb 2026.

TwIL-8B’s value is not “another leaderboard No. 1.” It puts a question on the table: if an 8.8B model can complete enough code, math, and instruction-following tasks, how much of the big-model bill is spent on memorizing knowledge? Karpathy’s five judgments will not all become true because of one model, and TwIL-8B’s knowledge weakness is plain to see. But local deployment, cost, and controllability now have a concrete model to point to.

Original tweets: mervenoyann and David Stout.