What fits your Mac.

1. guardian-3.1-2b-Q4_K_M.ggufno write-up

   Needs about 2.8 GB (estimate) · Loads on an 8 GB Mac or bigger

   q4_k_m · 1.5 GB file · llama.cpp · best 0.0% on screening

2. Gemma 4 E2BMeasured, not seated

   Needs about 3.3 GB (measured) · Loads on an 8 GB Mac or bigger

   The base model for the tagger fine-tune, and the smallest thing worth measuring.

   gemma-4-E2B_q4_0-it.gguf · q4_0 · 3.3 GB file · llama.cpp · mlx_lm · omlx · measured at 128K context: 4.1 GB

   Recorded as gemma4-e2b-b10618, gemma4-e2b-ab, gemma4-e2b-r1 and 2 more.

3. Ternary-Bonsai-8B-PQ2_0.ggufno write-up

   Needs about 3.5 GB (estimate) · Loads on an 8 GB Mac or bigger

   PQ2_0 · 2.2 GB file · llama.cpp · best 100.0% on toolscale

4. NeoHorse-1-4B-Q4_0.ggufno write-up

   Needs about 4.0 GB (estimate) · Loads on an 8 GB Mac or bigger

   Q4_0 · 2.5 GB file · llama.cpp · best 100.0% on toolscale

5. Qwen3.5-4B-Q4_0.ggufno write-up

   Needs about 4.0 GB (estimate) · Loads on an 8 GB Mac or bigger

   Q4_0 · 2.6 GB file · llama.cpp · best 100.0% on toolscale

6. Spark-X2.5-4B (XHToken)Measured, not seated

   Needs about 4.0 GB (estimate) · Loads on an 8 GB Mac or bigger

   Hybrid attention - one full-attention layer per three sliding-window layers - giving a native 1M context. Apache 2.0, 20T training tokens, marketed for agentic workflows with claimed native compatibility for Codex, Claude Code and Hermes.

   Spark-X2.5-4B-Q4_K_M.gguf · Q4_K_M · 2.6 GB file · llama.cpp

   Recorded as spark-x25-4b.

7. MiniCPM5-2B-Q8_0.ggufno write-up

   Needs about 4.1 GB (estimate) · Loads on an 8 GB Mac or bigger

   Q8_0 · 2.7 GB file · llama.cpp · best 100.0% on toolscale

   Also recorded as minicpm5-2b-q8-np1, minicpm5-2b-q8-nojinja, minicpm5-2b-q8-np4-today.

8. K2-Horizon-3.7B (IFM) · q4_0another build

   Needs about 4.5 GB (estimate) · Loads on an 8 GB Mac or bigger

   K2-Horizon-3.7B-Q4_0.gguf · q4_0 · 3.0 GB file · llama.cpp · best 100.0% on toolcall

   Also recorded as k2-3.7b-b11552-noreasonfmt, k2-3.7b-fork-control-nocoh.

9. K2-Horizon-3.7B (IFM)Seated · Dispatch

   Needs about 4.7 GB (estimate) · Loads on an 8 GB Mac or bigger

   The smallest model measured that perceives tool need. Held at three precisions (Q4/Q8/BF16) which made it the subject of the precision ladder.

   K2-Horizon-3.7B-Q4_K_M.gguf · Q4_K_M · 3.2 GB file · llama.cpp

   Recorded as k2-3.7b-q4-precision.

10. Gemma 4 E2B · mlx 4bitanother build

    Needs about 5.1 GB (estimate) · Loads on an 8 GB Mac or bigger

    gemma4-e2b-mlx-4bit · mlx 4bit · 3.6 GB file · mlx_lm · best 85.8% on tagger

11. Qwen3-8B · q4_0another build

    Needs about 6.6 GB (estimate) · Loads on a 16 GB Mac or bigger

    Qwen3-8B-Q4_0.gguf · q4_0 · 4.8 GB file · llama.cpp · best 47.2% on nullschema

12. Granite Guardian 4.1 8B · q4_0another build

    Needs about 6.6 GB (estimate) · Loads on a 16 GB Mac or bigger

    granite-guardian-4.1-8b-Q4_0.gguf · q4_0 · 4.8 GB file · llama.cpp · best 90.0% on screening

    Also recorded as granite-guardian-4.1-8b-b10893-control-nocoh.

13. Qwen3-8BMeasured, not seated

    Needs about 6.9 GB (estimate) · Loads on a 16 GB Mac or bigger

    The small reference and runtime test subject - the only model held in GGUF, MLX-native and BF16 simultaneously, which carried the whole runtime comparison. HumanEval 86.8% strict, beating qwen3.6-35b-a3b by being less verbose.

    Qwen3-8B-Q4_K_M.gguf · Q4_K_M · 5.0 GB file · llama.cpp

    Recorded as qwen3-8b-midrange.

14. Granite Guardian 4.1 8BSeated · Screen

    Needs about 7.0 GB (estimate) · Loads on a 16 GB Mac or bigger

    A purpose-built safety and screening classifier, tested for the judge seat where a general model was overkill.

    granite-guardian-4.1-8b-Q4_K_M.gguf · Q4_K_M GGUF, and MLX mixed_3_6 at 3.8GB · 5.1 GB file · llama.cpp · mlx_lm · omlx

    Recorded as granite-guardian-gguf-q4km.

15. gemma-4-E4B_q4_0-it.ggufno write-up

    Needs about 7.0 GB (estimate) · Loads on a 16 GB Mac or bigger

    q4_0 · 5.2 GB file · llama.cpp · best 100.0% on tools

    Also recorded as gemma4-e4b-b10618, gemma4-e4b-r1, gemma4-e4b-r2 and 1 more.

16. LFM 2.5 8B-A1BMeasured, not seated

    Needs about 7.0 GB (estimate) · Loads on a 16 GB Mac or bigger

    Tested as an extremely sparse small MoE, 1B active, to find the floor of useful.

    LFM2.5-8B-A1B-Q4_K_M.gguf · Q4_K_M · 5.2 GB file · llama.cpp

    Recorded as lfm25-8b-a1b-midrange.

17. K2-Horizon-3.7B (IFM) · q8_0another build

    Needs about 7.3 GB (estimate) · Loads on a 16 GB Mac or bigger

    K2-Horizon-3.7B-Q8_0.gguf · q8_0 · 5.4 GB file · llama.cpp · best 100.0% on screening

    Also recorded as k2-3.7b-q8-precision, k2-3.7b.

18. Ornith-1.5-9B-Q4_0.ggufno write-up

    Needs about 7.4 GB (estimate) · Loads on a 16 GB Mac or bigger

    q4_0 · 5.5 GB file · llama.cpp · best 100.0% on epistemics

    Also recorded as ornith-9b-kvq4, ornith-9b-kv-f16, ornith-9b-kv-q40.

19. K2-Horizon-7B-Q4_K_M.ggufno write-up

    Needs about 7.5 GB (estimate) · Loads on a 16 GB Mac or bigger

    q4_k_m · 5.6 GB file · llama.cpp · best 37.5% on nullschema

20. Qwen3.5-9B-Q4_K_M.ggufno write-up

    Needs about 7.6 GB (estimate) · Loads on a 16 GB Mac or bigger

    q4_k_m · 5.7 GB file · llama.cpp · best 23.6% on nullschema

21. Ternary-Bonsai-2-27B-PTQ1_0.ggufno write-up

    Needs about 7.9 GB (estimate) · Loads on a 16 GB Mac or bigger

    PTQ1_0 - 1.75 bpw ternary (group 128) · 5.9 GB file · llama.cpp · best 100.0% on epistemics

22. Gemma 4 12BSeated · Tag

    Needs about 8.1 GB (measured) · Loads on a 16 GB Mac or bigger

    Tested as a cheap tagger after the 80B proved expensive for a task that runs constantly.

    gemma-4-12b-it-qat-q4_0.gguf · q4_0 QAT · 7.0 GB file · llama.cpp · measured at 256K context: 12.1 GB

    Recorded as gemma4-12b-b10618, gemma4-12b-midrange, gemma4-12b-ab and 6 more.

23. Qwen3.8 27B · iq1_sanother build

    Needs about 8.2 GB (estimate) · Loads on a 16 GB Mac or bigger

    Qwen3.8-27B-UD-IQ1_S.gguf · iq1_s (unsloth UD) · 6.2 GB file · llama.cpp · best 77.1% on epistemics

24. Ternary-Bonsai-27B (PQ2_0)Measured, not seated

    Needs about 9.3 GB (estimate) · Loads on a 16 GB Mac or bigger

    prism-ml's group-128 ternary pack, built for their own llama.cpp fork's hybrid-attention kernels. The upstream Q2_g64 variant lacks them.

    Ternary-Bonsai-27B-PQ2_0.gguf · PQ2_0 (2.13 bpw, group-128) · 7.2 GB file · llama.cpp · mlx_lm

25. Ternary-Bonsai-2-27B-PQ2_0.ggufno write-up

    Needs about 9.4 GB (estimate) · Loads on a 16 GB Mac or bigger

    PQ2_0 - 2.13 bpw (group 128) · 7.2 GB file · llama.cpp · best 100.0% on epistemics

26. Ternary-Bonsai-27B (prism-ml)Measured, not seated

    Needs about 9.8 GB (estimate) · Loads on a 16 GB Mac or bigger

    End-to-end ternary weights on a Qwen3.6-27B backbone with ~75% linear attention, claiming 262K context and 94.6% of FP16 intelligence retained. 11.7 GB resident at a 128k window - a third of K2-36B for the same context.

    Ternary-Bonsai-27B-Q2_g64.gguf · Q2_g64 (1.71 bits/weight) · 7.6 GB file · llama.cpp

    Recorded as bonsai-27b-ternary.

27. Underdog-Saluki-27B-1.0-IQ2-mix.ggufno write-up

    Needs about 10.2 GB (estimate) · Loads on a 16 GB Mac or bigger

    IQ2_XS - 2.3125 bpw (IQ2-mix) · 7.9 GB file · llama.cpp · best 100.0% on tagger

28. Ternary-Bonsai-27B (prism-ml) · mlx 2bit g128another build

    Needs about 10.9 GB (estimate) · Loads on a 16 GB Mac or bigger

    ternary-bonsai-27b-mlx2 · mlx 2bit g128 · 8.5 GB file · mlx_lm · best 100.0% on toolcall

    Also recorded as ternary-bonsai-27b-recheck.

29. Ternary-Bonsai-27B (prism-ml) · mlx 2bit g128another build

    Needs about 10.9 GB (estimate) · Loads on a 16 GB Mac or bigger

    Ternary-Bonsai-27B-mlx-2bit (HF snapshot 70f75f3) · mlx 2bit g128 · 8.5 GB file · mlx_lm · best 100.0% on toolcall

30. Qwen3-8B · mlx 8bitanother build

    Needs about 11.1 GB (estimate) · Loads on a 16 GB Mac or bigger

    qwen3-8b-mlx-8bit · mlx 8bit · 8.7 GB file · mlx_lm · best 96.7% on screening

31. Qwen3.5-9B-Q8_0.ggufno write-up

    Needs about 12.1 GB (estimate) · Loads on an 18 GB Mac or bigger

    q8_0 · 9.5 GB file · llama.cpp · best 96.7% on screening

    Also recorded as qwen3.5-9b-q8.

32. K2-Horizon-7B-Q8_0.ggufno write-up

    Needs about 12.1 GB (estimate) · Loads on an 18 GB Mac or bigger

    q8_0 · 9.6 GB file · llama.cpp · best 100.0% on tools

    Also recorded as k2-7b, k2-7b-q8-precision.

33. Qwen3.8-9B-Distill (Q8_0)Measured, not seated

    Needs about 12.4 GB (estimate) · Loads on an 18 GB Mac or bigger

    Full-parameter distillation of Qwen3.8 2.4T A95B into the Qwen3.5-9B architecture, ~70k curated teacher traces. Tested as a cheap utility seat.

    Qwen3.8-9B-Q8_0.gguf · Q8_0 · 9.8 GB file · llama.cpp

    Recorded as qwen3.8-9b-distill-recheck.

34. K2-Horizon-3.7B (IFM) · bf16another build

    Needs about 12.8 GB (estimate) · Loads on an 18 GB Mac or bigger

    K2-Horizon-4B-BF16.gguf · bf16 · 10.1 GB file · llama.cpp · best 100.0% on toolcall

35. Qwen3.8 27B · mixedanother build

    Needs about 15.1 GB (estimate) · Loads on a 24 GB Mac or bigger

    qwen38-chad-default · mixed (mlx 3bit g64 base + 498 per-tensor overrides) · 12.1 GB file · mlx_lm · best 92.5% on worldview

36. gpt-oss-20b-MXFP4.ggufno write-up

    Needs about 15.1 GB (estimate) · Loads on a 24 GB Mac or bigger

    MXFP4 · 12.1 GB file · llama.cpp · best 86.2% on tagger

    Also recorded as gpt-oss-20b-study-default.

37. Qwen3.8-27B GSQ-RCO IQ3_S (ISTA-DASLab)Measured, not seated

    Needs about 15.1 GB (estimate) · Loads on a 24 GB Mac or bigger

    Same weights as qwen3.8-27b at ~3 bits instead of ~4.5, with the quantisation type chosen per tensor by gradient search under a size budget. Retains the MTP head. 26% fewer bytes per token and 8% faster.

    Qwen3.8-27B-GSQ-RCO-IQ3_S-mtp.gguf · IQ3_S (per-tensor gradient search) · 12.1 GB file · llama.cpp

    Recorded as qwen38-gsqrco-iq3s.

38. Devstral Small 2 24BMeasured, not seated

    Needs about 17.9 GB (estimate) · Loads on a 32 GB Mac or bigger

    A coding model, tested to see whether coding tuning transfers to structured extraction.

    Devstral-Small-2-24B-Instruct-2512-UD-Q4_K_XL.gguf · UD-Q4_K_XL · 14.5 GB file · llama.cpp

39. Muse-Glimmer 30BMeasured, not seated

    Needs about 18.0 GB (measured) · Loads on a 32 GB Mac or bigger

    Previously dismissed at 10.4 tok/s without ever being scored. Re-examined because it is multimodal, and scanned invoices are the part of the accountancy job nothing else can read.

    Muse-Glimmer-30B-UD-Q4_K_XL.gguf · UD-Q4_K_XL · 15.9 GB file · llama.cpp · measured at 128K context: 20.4 GB

40. Qwen3.8 27B · q4_0another build

    Needs about 19.3 GB (estimate) · Loads on a 32 GB Mac or bigger

    Qwen3.8-27B-Q4_0.gguf · q4_0 · 15.7 GB file · llama.cpp · best 100.0% on toolcall

41. Qwen3.8 27BSeated · agentic

    Needs about 20.2 GB (estimate) · Loads on a 32 GB Mac or bigger

    The dense counterweight to the MoE fleet, and the model where MLX first beat llama.cpp. Promoted to the agent seat on 2026-09-09 after toolcall/1.

    Qwen3.8-27B-UD-Q4_K_M.gguf · MLX 4-bit (MTP build, template patched to reasoning_effort=medium) · 16.5 GB file · llama.cpp

    Recorded as qwen38-27b-np1, qwen38-rebase-b10893, qwen38-vendor-qwen-think.

42. Granite 4.2 30B · mlx 4bitanother build

    Needs about 20.2 GB (estimate) · Loads on a 32 GB Mac or bigger

    granite-4.2-30b-mlx-q4 · mlx 4bit · 16.5 GB file · mlx_lm · best 91.7% on screening

43. diffusiongemma-26b-mlx-4bitno write-up

    Needs about 20.3 GB (estimate) · Loads on a 32 GB Mac or bigger

    4bit (MLX) · 16.5 GB file · mlx-vlm · best 100.0% on toolscale

44. GLM-4.7-Flash · Q4_0another build

    Needs about 21.0 GB (estimate) · Loads on a 32 GB Mac or bigger

    GLM-4.7-Flash-Q4_0.gguf · Q4_0 · 17.2 GB file · llama.cpp · best 100.0% on toolscale

45. Qwen3.8 27B · splash-packed-q4another build

    Needs about 21.2 GB (estimate) · Loads on a 32 GB Mac or bigger

    Qwen3.8-27B-Splash · splash-packed-q4 (4-bit g64) · 17.4 GB file · splash · best 100.0% on epistemics

    Also recorded as qwen3.8-27b-splash-nothink.

46. Granite 4.2 30BMeasured, not seated

    Needs about 21.6 GB (estimate) · Loads on a 32 GB Mac or bigger

    Tested because IBM builds explicitly for structured JSON output, and nullschema is the weakest axis across the whole fleet.

    granite-4.2-30b-Q4_K_M.gguf · Q4_K_M · 17.7 GB file · llama.cpp · mlx_lm · omlx

    Recorded as granite-4.2-30b.

47. K2-Horizon-7B-BF16.ggufno write-up

    Needs about 22.0 GB (estimate) · Loads on a 32 GB Mac or bigger

    bf16 · 18.0 GB file · llama.cpp · best 37.5% on nullschema

48. Qwen3 Coder 30BMeasured, not seated

    Needs about 22.6 GB (estimate) · Loads on a 32 GB Mac or bigger

    The first coding model measured against a real coding contract rather than an extraction suite.

    Qwen3-Coder-30B-A3B-Instruct-Q4_K_M.gguf · Q4_K_M · 18.6 GB file · llama.cpp

    Recorded as qwen3-coder-30b-b10618.

49. laguna-xs2-mlx-q4no write-up

    Needs about 22.9 GB (estimate) · Loads on a 32 GB Mac or bigger

    4bit (MLX) · 18.8 GB file · mlx_lm

50. Ornith 1.5 35B · mlx 4bitanother build

    Needs about 23.7 GB (estimate) · Loads on a 36 GB Mac or bigger

    ornith-1.5-35b-mlx-4bit · mlx 4bit · 19.5 GB file · mlx_lm · best 88.3% on screening

51. edge0-35bno write-up

    Needs about 23.9 GB (estimate) · Loads on a 36 GB Mac or bigger

    mlx 4bit g64 affine (MoE gates 8bit) · 19.7 GB file · edge0 · mlx_lm · best 97.9% on epistemics

    Also recorded as edge0-35b-recheck.

52. Qwen3.6 35B-A3BMeasured, not seated

    Needs about 24.8 GB (estimate) · Loads on a 36 GB Mac or bigger

    Tested as the fast general MoE, and used as the daily coding driver in DSH. Moved to an MLX build on 2026-09-09.

    Qwen3.6-35B-A3B-4bit · MLX 4-bit (also held as UD-Q5_K_M GGUF) · 20.4 GB file · rapid-mlx · swiftlm

    Recorded as mlx-community-Qwen3.6-35B-A3B-4bit, qwen3.6-35b-swiftlm, q36-toolparser.

53. Ornith 1.5 35B · q4_0another build

    Needs about 24.8 GB (estimate) · Loads on a 36 GB Mac or bigger

    Ornith-1.5-35B-A3B-Q4_0.gguf · q4_0 · 20.4 GB file · llama.cpp · best 100.0% on toolscale

    Also recorded as ornith-35b-a3b-q40-recheck, ornith-35b-tagger.

54. Qwen3.6 35B-A3B · splash-packed-q4-moeanother build

    Needs about 25.4 GB (estimate) · Loads on a 36 GB Mac or bigger

    Qwen3.6-35B-A3B-Splash · splash-packed-q4-moe (q4/q8, group 64) · 20.9 GB file · splash · best 100.0% on epistemics

55. Agents-A1 35BMeasured, not seated

    Needs about 25.6 GB (estimate) · Loads on a 36 GB Mac or bigger

    An agent-tuned MoE, tested on the tool suites where its specialism should show.

    Agents-A1-Q4_K_M.gguf · Q4_K_M · 21.2 GB file · llama.cpp

    Recorded as agents-a1-35b-b10618.

56. K2-Horizon-36B-A4B (IFM) · q4_0another build

    Needs about 25.8 GB (estimate) · Loads on a 36 GB Mac or bigger

    K2-Horizon-36B-A4B-Q4_0.gguf · q4_0 · 21.3 GB file · llama.cpp · best 100.0% on toolcall

57. BTL-4 (Q4_K_M)Measured, not seated

    Needs about 26.1 GB (estimate) · Loads on a 36 GB Mac or bigger

    Fine-tuned from Ornith-1.0-35B on an execution-gated corpus: trajectories kept only where the code ran and passed. 256 experts, 8 routed, 262k native context, ~20 KB/token KV.

    BTL-4-Q4_K_M.gguf · Q4_K_M (community conversion) · 21.5 GB file · llama.cpp

    Recorded as btl-4-q4km-recheck.

58. Qwen-AgentWorld-35B-A3B-UD-Q4_K_M.ggufno write-up

    Needs about 26.8 GB (estimate) · Loads on a 48 GB Mac or bigger

    UD-Q4_K_M · 22.1 GB file · llama.cpp · best 80.0% on tools

59. K2-Horizon-36B-A4B (IFM)Seated · agent

    Needs about 27.0 GB (estimate) · Loads on a 48 GB Mac or bigger

    MoVA (Mixture-of-Value Attention) extends sparsity into the value component of attention. 4B active gives it 25 tok/s at 5k where a dense 27B manages 8.6. Ties the top of the board on tool invocation and holds a 200-step chain without deviating.

    K2-Horizon-36B-A4B-Q4_K_M.gguf · Q4_K_M · 22.4 GB file · llama.cpp

    Recorded as k2-36b-ctx32k, k2-36b-a4b-q4, k2-36b-extraction and 4 more.

60. Qwen3.6 35B-A3B · Q5_K_Manother build

    Needs about 32.5 GB (estimate) · Loads on a 48 GB Mac or bigger

    Qwen3.6-35B-A3B-UD-Q5_K_M.gguf · Q5_K_M · 27.1 GB file · llama.cpp · omlx · rapid-mlx · splash · swiftlm · best 100.0% on epistemics

    Also recorded as qwen3.6-35b-ag, qwen3.6-35b-tag.

61. Ornith 1.5 35BMeasured, not seated

    Needs about 35.0 GB (estimate) · Loads on a 48 GB Mac or bigger

    A second non-Qwen coding candidate, tested alongside Devstral.

    Ornith-1.5-35B-Q6_K.gguf · Q6_K · 29.2 GB file · llama.cpp · mlx_lm

62. Sakura-Kolibri-1-365E-IQ3_XXS-27.45GiB.ggufno write-up

    Needs about 35.3 GB (estimate) · Loads on a 48 GB Mac or bigger

    IQ3_XXS mix (3.17 bpw) · 29.5 GB file · llama.cpp · best 100.0% on epistemics

    Also recorded as sakura-iq3xxs-np2-128k-b11533.

63. GLM-4.7-FlashMeasured, not seated

    Needs about 38.1 GB (estimate) · Loads on a 64 GB Mac or bigger

    Tested as a fast MoE with a 202k context window, and as the first model we quantised ourselves for MLX.

    GLM-4.7-Flash-Q8_0.gguf · Q8_0 · 31.8 GB file · llama.cpp · mlx_lm

64. Qwen3-Next-80B-A3B Instruct, REAP prunedMeasured, not seated

    Needs about 39.0 GB (estimate) · Loads on a 64 GB Mac or bigger

    Tested to find out whether expert pruning costs the extraction seat its abstention. The full model holds that seat at 40 GB, which is most of the memory budget on a 64 GB machine.

    Qwen3-Next-80B-A3B-Instruct-REAP.IQ4_XS.gguf · IQ4_XS · 32.7 GB file · llama.cpp · mlx_lm · omlx · rapid-mlx

    Recorded as qwen3-next-80b-reap-ag.

65. Qwen3-Next-80B-A3B Instruct, REAP pruned · Q4_K_Manother build

    Needs about 43.8 GB (estimate) · Loads on a 64 GB Mac or bigger

    Qwen3-Next-80B-A3B-Instruct-REAP.Q4_K_M.gguf · Q4_K_M · 36.7 GB file · llama.cpp · best 87.5% on toolscale

66. K2-Horizon-36B-A4B (IFM) · q8_0another build

    Needs about 47.4 GB (estimate) · Loads on a 64 GB Mac or bigger

    K2-Horizon-36B-A4B-Q8_0.gguf · q8_0 · 39.8 GB file · llama.cpp · best 70.8% on nullschema

67. Qwen3-Coder-Next-IQ4_XS.ggufno write-up

    Needs about 50.7 GB (estimate) · Loads on a 64 GB Mac or bigger

    IQ4_XS · 42.7 GB file · llama.cpp · best 90.0% on screening

68. Kolibri-1-Q4_0.ggufno write-up

    Needs about 52.4 GB (estimate) · Loads on a 96 GB Mac or bigger

    Q4_0 · 44.2 GB file · llama.cpp · best 100.0% on toolcall

macOS doesn’t hand all of a Mac’s memory to the graphics chip that runs the model. It keeps the rest for itself and your apps. This is how much it gave on each of our Macs, with nothing changed.