rtx-5090-benchmarks / reports /lfm25-24b-vs-qwen36-claims.md
witcheer's picture
report: LFM2.5-24B-A2B vs Qwen3.6-35B-A3B card-claim check
d748f4f verified
|
Raw History Blame Contribute Delete
5.41 kB

LFM2.5-24B-A2B against Qwen3.6-35B-A3B: checking the model card's two claims on one RTX 5090

The one line: Liquid's model card says LFM2.5-24B-A2B beats Qwen3.6-35B-A3B on IFEval and trails it on GPQA Diamond. On the same items, thinking on, a 16k budget and 5-bit GGUFs, both directions hold: IFEval 93.35 vs 87.25, GPQA Diamond 68.69 vs 76.77. Most of the GPQA gap is the budget: 48 of LFM's 62 GPQA misses reached the 16,384-token cap mid-thought, against 13 of Qwen's 46.

Rig: RTX 5090 32GB (capsule), llama-server, batch 1 · Window: 2026-10-14 to 2026-10-15, inside the box's cheap-electricity hours (idle-queue item 65t-lfm25-24b) · Harness: llm-bench-rig · Status: partial. IFEval (both LFM files) and GPQA Diamond (5-bit file) are finished; MATH-500, HumanEval+ and MBPP+ on both files and the 8-bit GPQA leg are still queued. This report covers only the finished legs, which are the two tasks the card names against Qwen3.6-35B-A3B.

LFM2.5-24B-A2B vs Qwen3.6-35B-A3B

TL;DR

  1. IFEval: the card's direction holds, by more. LFM2.5-24B-A2B UD-Q5_K_XL 93.35 (505/541, prompt-strict) vs Qwen3.6-35B-A3B UD-Q5_K_M 87.25 (472/541). Paired on items: LFM right where Qwen is wrong 39 times, the reverse 6 times (exact two-sided binomial p < 0.0001). LFM also thinks shorter here: median 1,660 completion tokens vs 2,375.
  2. GPQA Diamond: the card's direction holds too. 68.69 (136/198) vs 76.77 (152/198). Paired: Qwen-only right 29, LFM-only right 13 (p = 0.020).
  3. The GPQA gap is mostly the 16k budget. LFM's median GPQA completion is 6,992 tokens vs 2,981 for Qwen. 48 of LFM's 62 misses reached the cap; finished-and-wrong answers are LFM 14, Qwen 33. The card's own evaluation used a 32,768-token cap. A 32k GPQA leg on both rows has not been run.
  4. Quant tier barely moves IFEval. The 8-bit LFM file scores 92.61 (501/541), 0.74 under the 5-bit file.
  5. Decode speed. llama-bench tg128 on this card: LFM UD-Q5_K_XL 444.0 tok/s (b10371, 2026-10-14) vs the board's Qwen3.6-35B-A3B UD-Q4_K_M 270.97 tok/s and UD-Q6_K 259.65 tok/s (older build, 2026-05-28, dataset/benchmarks.csv). Served decode during the LFM eval: 378 tok/s, single stream.

Setup

  • Regime, every row: thinking on, max_tokens 16384, ctx 24576, greedy (temperature 0), zero-shot, batch 1 through llama-server's chat-completions endpoint.
  • Tasks: IFEval (541 prompts, prompt-strict; language:response_language is skipped as unsupported, same for both rows), GPQA Diamond (198 questions, shuffle seed 42).
  • Subject: LiquidAI/LFM2.5-24B-A2B as GGUF, UD-Q5_K_XL and Q8_0 (unsloth/LFM2.5-24B-A2B-GGUF). Licence: Liquid's own lfm1.0, not Apache.
  • Comparison row (already on the second-tier 16k board, not re-run): Qwen3.6-35B-A3B UD-Q5_K_M, 26,456,194,016 bytes, dataset/second_tier.csv.
  • Server binary: the LFM legs ran on the rig's mainline llama-server, b10371 (5d16e81dd). The Qwen row predates per-row build logging, so the same build is not claimed.
  • Paired comparisons: per-item correct and capped flags from each task's *_tokens.jsonl, joined on item index; exact two-sided binomial on the discordant items.

The two claims

Task LFM2.5-24B-A2B UD-Q5_K_XL Qwen3.6-35B-A3B UD-Q5_K_M Card: LFM Card: Qwen Paired (LFM-only / Qwen-only)
IFEval 93.35 (505/541) 87.25 (472/541) 96.75 93.06 39 / 6, p < 0.0001
GPQA Diamond 68.69 (136/198) 76.77 (152/198) 74.42 83.69 13 / 29, p = 0.020

The card's numbers come from vLLM at temperature 0.6 with a 32,768-token cap ("Other evals" in the card's evaluation settings), so both models score higher there; this board is greedy, 16k, 5-bit GGUF.

Where the misses come from

Task Row Misses Capped misses Finished and wrong Median completion tokens
IFEval LFM Q5 36 0 36 1,660
IFEval Qwen Q5 69 17 52 2,375
GPQA Diamond LFM Q5 62 48 14 6,992
GPQA Diamond Qwen Q5 46 13 33 2,981

Limits

  • One pass per item, greedy; GPQA's 198 items put one item at ~0.5 points.
  • The finished-and-wrong count says LFM's GPQA losses are mostly budget, not that a bigger budget would turn them into right answers. That needs the 32k leg.
  • Tool calling is not tested. The card says the model was trained with agentic RL inside agent harnesses (OpenClaw, Hermes Agent, Pi, OpenCode). On llama.cpp, ggml-org/llama.cpp#30490 (open at writing) reports its tool calls arriving as plain-text XML and being lost, so check tool-call parsing before agent use.
  • The speed comparison crosses builds and quant tiers (LFM 5-bit on b10371 vs Qwen 4-bit and 6-bit on a May build).

Sources

  • capsule ~/benchmark-rig/results/lfm2-5-24b-a2b-ud-q5-k-xl-thinkon/{ifeval_detail.json,gpqa.json,ifeval_tokens.jsonl,gpqa_tokens.jsonl}, lfm2-5-24b-a2b-q8-0-thinkon/ifeval_detail.json, qwen3-6-35b-a3b-ud-q5-k-m-thinkon/{ifeval_detail.json,gpqa.json,*_tokens.jsonl}, read 2026-10-15.
  • Model card https://huggingface.co/LiquidAI/LFM2.5-24B-A2B, read 2026-10-15.
  • Chart: scripts/chart_lfm25_claims.py.