Calibrating 2-bit GGUFs (<10Gb) for agentic coding tasks

TL;DR: Small quantizations (< 10 Gb) of Qwopus3.6-27B-Coder calibrated on agentic coding logs with a bundled MTP that achieve >60% pass rate on SWE-rebench. What's included: 📦 3 imatrix-calibrated quants: IQ2_XS (8.9 GiB), IQ2_M (9.7 GiB), Q2_K_S (9.96 GiB) ⚡ MTP draft head kept lossless at Q8 while trunk goes 2-bit → 1.26× decode speedup (79.9% acceptance, n-max=1) 🎯 Calibrated on real agentic-coding logs (Claude Code, Qwen Code, opencode; English + Python focused) 🔬 Hybrid importance matrix (activation + weight energy) with special-token parsing to protect tool-call channels The IQ2_M quant achieves a strong 63% pass rate on the nebius/SWE-rebench agentic coding benchmark which is comparable to the pass rate of the Q5_K_M quant, despite being half the size. The IQ2_M quant is also more robust to loops than a non-calibrated quant of the same stature but not as robust as the Q5_K_M unless the repetition penalty is set to >1. Metrics Metric FP16 (reference) Q2_K IQ2_XS IQ2_M Q2_K_S File n/a Q2_K.gguf IQ2_XS.gguf IQ2_M.gguf Q2_K_S.gguf Quality ❌ ❌ ⭐⭐⭐ ⭐⭐ Technique none none imatrix imatrix imatrix Size (GiB) 50.90 10.40 8.89 9.74 9.96 BPW 16.000 3.269 2.794 3.062 3.133 PPL (general) 6.4826 5.5835 9.8866 8.5961 8.0091 KLD med (general) 0.00000 0.1154 0.0950 0.0535 0.0566 top_p (general) 100.00% 79.29% 78.87% 83.23% 83.32% Plain Q2_K scores worse KLD than calibrated IQ2_M despite being larger i.e. the calibration matters. SWE-rebench Results The agentic coding capabilities of each quant were evaluated on 10 real-world coding issues from the nebius/SWE-rebench using the OpenAI Agents SDK pointed at a local llama-server. For each nebius/SWE-rebench issue, the agent gets the problem statement and a live bash tool that shells into a dedicated Docker container with the repo pre-checked out at the failing commit. It iterates by reading files, running tests, editing code until it produces a git diff or hits the step limit. The patch is then graded by actually running the repo's FAIL_TO_PASS test suite inside the container, so pass/fail is real execution, not fuzzy matching. We tried using mini SWE-Agent but it wasn't adequately resolving issues despite have a similar patch rate. Metric Q2_K IQ2_XS IQ2_M Q2_K_S Q5_K_M File Q2_K.gguf IQ2_XS.gguf IQ2_M.gguf Q2_K_S.gguf Q5_K_M.gguf Technique none imatrix imatrix imatrix none Size (GiB) 10.40 8.89 9.74 9.96 19.50 Repetitions 3 3 3 3 3 Issues 10 10 10 10 10 Patch Rate 88±12% 70±10% 100% 93±6% 100% Pass Rate 30±10% 27±6% 63±6% 57±6% 57±6% Max Turns 27±15% 57±25% 13±15% 10±17% 0% Mean Steps 58.5±7.6 73.1±15.1 51.6±8.3 46.7±8.1 38.6±1.3 Mean Tokens 1,335K±253K 1,779K±137K 784K±260K 922K±195K 588K±57K Tool Error Rate 14.6±6.4% 9.5±3.6% 12.6±1.8% 8.9±1.5% 12.1±0.2% Mean Wall 415±98s 558±182s 381±66s 425±259s 307±34s Sampling Parameters: temperature=0.25, top_p=0.95, top_k=20, min_p=0.0, presence_penalty=0.0, repetition_penalty=1.0, max_tokens=32768, ctx=131072, thinking=true, mtp=true, mtp_draft_n_max=2. Tested on 4060Ti (16Gb) Definitions: patched - how many of the 10 issues did the agent produce a patch for (even if it didn't resolve)? resolved - how many of the 10 issues had patches that passed all FAIL_TO_PASS tests? max_turns - how many of the 10 issues hit the 100-step cap without resolving? mean_steps - average number of agentic steps taken (shelling into Docker, reading files,editing code counts as steps) mean_tokens - average number of tokens generated across the entire agentic episode tool_err_rate - how often the agent produced an invalid shell command that couldn't be executed (syntax errors, wrong file paths, etc.) mean_wall - average wall-clock time per episode (capped at 2 hours for those that hit the step limit) Overall, the IQ2_M quant achieves a strong 63% pass rate on this agentic coding benchmark, which is impressive for a 2-bit model. The high patch rate across all quants suggests that even the weaker ones can still generate plausible patches, but the lower pass rates and higher max turn rates indicate that many of those patches aren't actually resolving the issues. The IQ2_M quant behaves as good as the Q5_K_M albiet with ~20% more steps and tokens, however those additional steps and iterations look to be effective ones that are helping it self-correct and resolve more issues, rather than just looping. When the quant has a high number of mean tokens in combination with a high max turn rate that usually indicates the agent is stuck in a loop. It's worth pointing out that Q5KM never hits its max turn (100) when solving these issues. We recommend running these quants with a repetition penalty of >1 to break it out of loops. Given the variation induced from sampling, we run a few repetitions of each quant and report the mean ± standard deviation across those runs. Quick start: ollama run hf.co/pearsonkyle/Qwopus3.6-27B-Coder-2bit-MTP-GGUF:IQ2_M For GPU with MTP speculative decoding: llama-server --model Qwopus3.6-27B-Coder-IQ2_M.gguf \ --spec-type draft-mtp --spec-draft-n-max 1 \ --flash-attn on --n-gpu-layers 999 Caveats: Sub-3.2-bpw quants — great when VRAM is the constraint, not a replacement for Q4+ when it is available Calibration was English + Python-heavy; expect weaker fidelity on other languages and non-coding workloads 📎 HF Repo · Quant-Tuner · Log Miner · Agent Source · Calibration Data