TwIL-8B — 8.8B Reasoning Model Post-Trained from Granite 4.2 with SFT + GRPO/DAPO RL

TL;DR

TwIL-8B is an 8.8B reasoning model from webAI, post-trained from IBM’s granite-4.2-8b base via full-parameter SFT followed by GRPO/DAPO reinforcement learning on math/code/STEM/instruction-following prompts. It ships in two variants that share the same architecture and tokenizer: balanced (shorter traces) and verbose (longer traces, higher accuracy on hard contests). Against open peers up to ~3.5× its size it leads on MATH-500 (98.2), LiveCodeBench (81.6), IFBench (74.4 vs 64.5 for Qwen3.5-9B), and reaches 88–93% of frontier-model scores on reasoning benchmarks — while using ~38% fewer tokens per AIME 2026 problem than Qwen3.5-9B. It runs in 4.98 GiB at Q4_K_M. The pitch is explicit: intelligence-per-dollar matters more than peak capability, and an 8B model on local hardware can do most of the work a frontier model does on reasoning-heavy tasks.

Key claims

Full-parameter SFT on reasoning traces for math/STEM/instruction-following, then GRPO/DAPO RL on the same domains plus code; a third length-aware RL stage produces the balanced variant by rewarding shorter traces against per-domain length references [§How it was built].
MATH-500 98.2 beats Qwen3.5-9B (96.6) and Gemma 4 31B (87.9); AIME 2025 88.3 beats Gemma 4 31B (87.1); LiveCodeBench 81.6 beats Gemma 4 31B (80.0), Gemma 4 12B (72.0), and Qwen3.5-9B (65.6) by 16 points [§Against open peers].
Instruction-following: IFBench 74.4 is +9.9 over Qwen3.5-9B, level with Gemma 4 12B (74.0), and beats Claude Opus 4.8 (62.2) by 12.2 points; IFEval 91.8 is level with Qwen3.5-9B (91.5) [§Highlights].
Token efficiency vs same-size open peer: ~19.5k output tokens per AIME 2026 problem vs 31.2k for Qwen3.5-9B (38% fewer) and 32.9k vs 40.6k on HMMT Feb 2026 (19% fewer) — at 4.2 and 3.0 fewer accuracy points respectively [§Tokens per problem].
Reaches 88–93% of the best frontier score (GPT-5.5 / Claude Opus 4.8 / Grok 4.3) on LiveCodeBench, IFBench, HMMT Nov 2025, AIME 2026, HMMT Feb 2025 — but only 69% on HMMT Feb 2026 and 26% on HLE [§Against frontier models].
Explicit non-goal: not a knowledge model. GPQA Diamond 70.2 trails Qwen3.5-9B by 11.5 and Gemma 4 31B by 14.1; MMLU-Pro and HLE show the same pattern; authors attribute this to knowledge scaling with parameter count more than with reasoning-focused training [§Limitations and caveats, §Against frontier models].
Ships bf16 safetensors (16.38 GiB) plus Q4_K_M GGUFs (4.98 GiB) for both variants; chat template opens a <think> block by default with temperature=1.0, top_p=0.95, and recommends a ≥32k-token generation budget [§Usage].
Training pools (math/STEM/instruction-following) were decontaminated with 13-gram matching against AIME, HMMT, MATH-500, GPQA Diamond, MMLU-Pro, HLE, IFEval, IFBench and LiveCodeBench; HMMT Nov 2025 and the HumanEval/MBPP/CRUXEval suites were NOT in the decontamination set [§Limitations and caveats].

Method

Base is ibm-granite/granite-4.2-8b — a GraniteForCausalLM with 40 layers, hidden size 4096, 32 attention heads / 8 KV heads, 8.79B parameters. Post-training has three stages: (1) full-parameter SFT on reasoning traces covering math, STEM and instruction following; (2) GRPO/DAPO RL on math/code/STEM/instruction-following prompts — this produces the verbose variant; (3) a further RL stage with a length-aware reward measured against per-domain length references — this produces the balanced variant. Both variants are shipped as full merged models (not LoRA adapters), with identical architecture, tokenizer, chat template, and generation config; only the amount of reasoning before answering differs. The chat template defaults to opening a <think>...</think> block before the answer and accepts enable_thinking=False / reasoning_effort="low" through chat_template_kwargs (with substantially lower scores at low effort — not used for any card number).

Results

Open-peer comparison (balanced variant unless marked †):

MATH-500: 98.2 (TwIL) vs 96.6 (Qwen3.5-9B) vs 87.9 (Gemma 4 31B).
AIME 2026: 88.3 vs 92.5 (Qwen3.5-9B) vs 77.5 (Gemma 4 12B) vs 89.2 (Gemma 4 31B).
HMMT Feb 2026: 68.2 vs 71.2 (Qwen3.5-9B) vs 77.2 (Gemma 4 31B) — the weakest row.
HMMT Nov 2025: 86.7 vs 82.9 vs 87.5.
LiveCodeBench †: 81.6 vs 65.6 (Qwen3.5-9B, +16.0) vs 72.0 (Gemma 4 12B) vs 80.0 (Gemma 4 31B).
IFBench †: 74.4 vs 64.5 (Qwen3.5-9B, +9.9) vs 74.0 (Gemma 4 12B) vs 76.0 (Gemma 4 31B).
GPQA Diamond: 70.2 vs 81.7 (Qwen3.5-9B, −11.5) vs 84.3 (Gemma 4 31B, −14.1) — the knowledge gap.

Share-of-frontier on key reasoning benchmarks: 93% on LiveCodeBench, 92% on IFBench, 90% on HMMT Nov 2025, 88% on AIME 2026, 88% on HMMT Feb 2025, 85% on MMLU-Pro, 75% on GPQA Diamond, 69% on HMMT Feb 2026, 26% on HLE.

Head-to-head records: 7–6 vs Qwen3.5-9B across 13 shared benchmarks (shared average 78.1 vs 77.6); 3–11 vs Gemma 4 31B (which has 3.5× the parameters); 0–9 vs GPT-5.5 shared average 71.5 vs 86.7.

Why it’s interesting

A clean data point for the “small models eat the middle” thesis: a focused SFT+GRPO/DAPO post-train on an 8B base reaches near-parity with a 31B dense model on code and instruction-following, and matches Qwen3.5-9B on contest math while spending ~20–40% fewer output tokens. The length-aware RL stage producing the balanced variant is the operationally interesting piece — it’s an explicit knob for the reasoning-budget tradeoff that complements the broader Inference-Time Scaling literature, where most work scales tokens up; here the reward targets fewer tokens at equal accuracy. The honest carve-out on knowledge benchmarks (GPQA/MMLU-Pro/HLE trailing by 10–30 points) reinforces what Scaling Laws for Neural Language Models and recent work on Internal Data Repetition Destroys Language Models imply: knowledge scales with parameters and data, reasoning scales with targeted training. Complements LoRA Without Regret‘s observations on how much heavy-tail reasoning capability can be injected into a fixed base via post-training alone.